| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Kimi-K3 decode GEMM selection for unquantized BF16 on SM90/SM100/SM103/SM107. |
| 4 | |
| 5 | Dispatch is purely by local ``(N, K)`` shape and token count ``M`` — the module |
| 6 | name plays no role. Each measured shape maps to a :class:`ProjectionSpec` |
| 7 | holding the winning backend per token count. The static part of the decision is |
| 8 | resolved once per module at install time into a small ``{M: call}`` plan, so the |
| 9 | per-forward path is a single dict lookup. |
| 10 | |
| 11 | The supported capabilities carry separate measured tables: |
| 12 | :data:`KIMI_K3_PROJECTIONS` was tuned on B300 (SM103), |
| 13 | :data:`KIMI_K3_PROJECTIONS_SM100` on B200 (SM100), and |
| 14 | :data:`KIMI_K3_PROJECTIONS_SM90` on H200 (SM90). The per-(shape, M) winners |
| 15 | genuinely differ between the parts, so the tables must not be merged. SM107 |
| 16 | (Rubin) reuses the SM103 table: the plan was validated end-to-end on SM107 |
| 17 | hardware, but the per-M crossovers have not been re-measured there and may |
| 18 | deserve their own table once retuned. |
| 19 | """ |
| 20 | |
| 21 | from __future__ import annotations |
| 22 | |
| 23 | from dataclasses import dataclass |
| 24 | from typing import Literal |
| 25 | |
| 26 | import torch |
| 27 | from torch import nn |
| 28 | |
| 29 | import vllm.envs as envs |
| 30 | from vllm import _custom_ops as ops |
| 31 | from vllm.model_executor.kernels.linear.cute_dsl.skinny_gemm import ( |
| 32 | SkinnyGemmConfig, |
| 33 | shape_dynamic_skinny_gemm, |
| 34 | ) |
| 35 | from vllm.model_executor.layers.linear import LinearBase, UnquantizedLinearMethod |
| 36 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 37 | ParallelLMHead, |
| 38 | UnquantizedEmbeddingMethod, |
| 39 | ) |
| 40 | from vllm.platforms import current_platform |
| 41 | from vllm.utils.multi_stream_utils import maybe_execute_in_parallel |
| 42 | |
| 43 | Backend = Literal["cute", "dsv3_fused_a"] |
| 44 | # A resolved per-token-count call: the backend plus its CuTe config (None for |
| 45 | # dsv3, which needs no config). |
| 46 | ResolvedCall = tuple[Backend, SkinnyGemmConfig | None] |
| 47 | |
| 48 | # TP8 KDA projection split, measured together under CUDA graph capture on B300. |
| 49 | # Q/K/V/G stay on the graph's main stream while F_A/beta and F_B run on the |
| 50 | # model's auxiliary stream. |
| 51 | KDA_M1_QKVG_CONFIG = SkinnyGemmConfig(1, 64, 4, 2, 8) |
| 52 | KDA_M1_FAB_CONFIG = SkinnyGemmConfig(1, 224, 1, 2, 8) |
| 53 | KDA_QKVG_CONFIGS = { |
| 54 | 1: KDA_M1_QKVG_CONFIG, |
| 55 | 2: SkinnyGemmConfig(2, 64, 3, 2, 8), |
| 56 | } |
| 57 | # The captured end-to-end projection sweep wins through M=14 and regresses at |
| 58 | # M=15 and M=16, where the original packed projection remains selected. |
| 59 | KDA_PROJECTION_OVERLAP_MAX_TOKENS = 14 |
| 60 | KDA_SKINNY_N_MAX_TOKENS = KDA_PROJECTION_OVERLAP_MAX_TOKENS |
| 61 | KDA_SKINNY_K_MAX_TOKENS = KDA_PROJECTION_OVERLAP_MAX_TOKENS |
| 62 | _KDA_QKVG_SIZE = 4 * 1536 |
| 63 | _KDA_FAB_SIZE = 128 + 12 |
| 64 | _KDA_PACKED_SIZE = 6288 |
| 65 | _KDA_TP_SIZE = 8 |
| 66 | |
| 67 | |
| 68 | def _kda_qkvg_flashinfer_backend() -> Literal["cute-dsl"] | None: |
| 69 | """Return the tested FlashInfer backend for the low-M QKVG branch. |
| 70 | |
| 71 | FlashInfer's ``cute-dsl`` BF16 GEMM supports SM100 and SM103, but rejects |
| 72 | SM90 during backend validation. Other capabilities use ``torch.mm`` for |
| 73 | QKVG while retaining the concurrent F_A/beta and F_B branch. |
| 74 | """ |
| 75 | if current_platform.is_device_capability( |
| 76 | (10, 0) |
| 77 | ) or current_platform.is_device_capability((10, 3)): |
| 78 | return "cute-dsl" |
| 79 | return None |
| 80 | |
| 81 | |
| 82 | @dataclass(frozen=True, slots=True) |
| 83 | class ProjectionSpec: |
| 84 | n: int |
| 85 | k: int |
| 86 | dsv3_tokens: frozenset[int] = frozenset() |
| 87 | cute_configs: tuple[tuple[int, SkinnyGemmConfig], ...] = () |
| 88 | residual_configs: tuple[tuple[int, SkinnyGemmConfig], ...] = () |
| 89 | name: str = "" # optional debug label; never used for dispatch |
| 90 | |
| 91 | def cute_config(self, num_tokens: int) -> SkinnyGemmConfig | None: |
| 92 | return dict(self.cute_configs).get(num_tokens) |
| 93 | |
| 94 | def residual_config(self, num_tokens: int) -> SkinnyGemmConfig | None: |
| 95 | return dict(self.residual_configs).get(num_tokens) |
| 96 | |
| 97 | |
| 98 | def _cute( |
| 99 | num_tokens: int, |
| 100 | block_size: int, |
| 101 | outputs_per_block: int, |
| 102 | k_unroll: int, |
| 103 | vector_width: int = 8, |
| 104 | ) -> SkinnyGemmConfig: |
| 105 | return SkinnyGemmConfig( |
| 106 | num_tokens, |
| 107 | block_size, |
| 108 | outputs_per_block, |
| 109 | k_unroll, |
| 110 | vector_width, |
| 111 | ) |
| 112 | |
| 113 | |
| 114 | _M1_TO_16 = frozenset(range(1, 17)) |
| 115 | _M1 = frozenset({1}) |
| 116 | |
| 117 | # Keyed by local (N, K). Where two projections share a shape (only 1536x7168: |
| 118 | # shared_gate_up_proj and mla_g_proj) the entry is unified. |
| 119 | KIMI_K3_PROJECTIONS: dict[tuple[int, int], ProjectionSpec] = { |
| 120 | (1536, 128): ProjectionSpec(1536, 128, _M1_TO_16, name="f_b_proj"), |
| 121 | (3072, 128): ProjectionSpec(3072, 128, _M1_TO_16, name="f_b_proj"), |
| 122 | # 1536x7168 is shared by shared_gate_up_proj and mla_g_proj. dsv3 M1..16 is |
| 123 | # only crash-safe once the mla_g aux-stream/PDL capture fix lands (subtask |
| 124 | # task_7388aba1); the fallback if it cannot be fixed is dsv3_tokens=_M1. |
| 125 | (1536, 7168): ProjectionSpec( |
| 126 | 1536, 7168, _M1_TO_16, name="shared_gate_up_proj/mla_g_proj" |
| 127 | ), |
| 128 | (3072, 7168): ProjectionSpec( |
| 129 | 3072, |
| 130 | 7168, |
| 131 | cute_configs=( |
| 132 | (1, _cute(1, 224, 3, 4)), |
| 133 | (2, _cute(2, 128, 3, 2)), |
| 134 | (3, _cute(3, 128, 2, 1)), |
| 135 | (4, _cute(4, 64, 2, 2)), |
| 136 | (5, _cute(5, 128, 3, 1)), |
| 137 | ), |
| 138 | name="shared_gate_up_proj", |
| 139 | ), |
| 140 | (2112, 7168): ProjectionSpec(2112, 7168, _M1_TO_16, name="fused_qkv_a_proj"), |
| 141 | (2304, 1536): ProjectionSpec(2304, 1536, _M1_TO_16, name="q_b_proj"), |
| 142 | (4608, 1536): ProjectionSpec(4608, 1536, _M1_TO_16, name="q_b_proj"), |
| 143 | (3584, 7168): ProjectionSpec( |
| 144 | 3584, |
| 145 | 7168, |
| 146 | frozenset(range(2, 9)), |
| 147 | ((1, _cute(1, 224, 2, 4)),), |
| 148 | name="routed_expert_down_proj", |
| 149 | ), |
| 150 | (6288, 7168): ProjectionSpec( |
| 151 | 6288, |
| 152 | 7168, |
| 153 | cute_configs=( |
| 154 | (1, _cute(1, 224, 3, 4)), |
| 155 | (2, _cute(2, 64, 3, 2)), |
| 156 | (3, _cute(3, 32, 3, 4)), |
| 157 | (4, _cute(4, 128, 6, 1)), |
| 158 | ), |
| 159 | name="in_proj_qkvgfab", |
| 160 | ), |
| 161 | (12448, 7168): ProjectionSpec( |
| 162 | 12448, |
| 163 | 7168, |
| 164 | cute_configs=( |
| 165 | (1, _cute(1, 224, 4, 2)), |
| 166 | (2, _cute(2, 64, 4, 2)), |
| 167 | (3, _cute(3, 64, 2, 2)), |
| 168 | ), |
| 169 | name="in_proj_qkvgfab", |
| 170 | ), |
| 171 | (7168, 768): ProjectionSpec(7168, 768, _M1_TO_16, name="shared_down_proj"), |
| 172 | (7168, 1536): ProjectionSpec( |
| 173 | 7168, 1536, cute_configs=((1, _cute(1, 96, 4, 2)),), name="o_proj" |
| 174 | ), |
| 175 | (7168, 3072): ProjectionSpec( |
| 176 | 7168, |
| 177 | 3072, |
| 178 | cute_configs=( |
| 179 | (1, _cute(1, 96, 2, 4)), |
| 180 | (2, _cute(2, 32, 4, 4)), |
| 181 | ), |
| 182 | name="o_proj", |
| 183 | ), |
| 184 | (7168, 3584): ProjectionSpec( |
| 185 | 7168, |
| 186 | 3584, |
| 187 | cute_configs=( |
| 188 | (1, _cute(1, 224, 4, 2)), |
| 189 | (2, _cute(2, 64, 4, 2)), |
| 190 | ), |
| 191 | residual_configs=( |
| 192 | (1, _cute(1, 64, 4, 2)), |
| 193 | (2, _cute(2, 64, 7, 2)), |
| 194 | (3, _cute(3, 64, 2, 1)), |
| 195 | (4, _cute(4, 64, 2, 1)), |
| 196 | ), |
| 197 | name="routed_expert_up_proj", |
| 198 | ), |
| 199 | (7168, 4224): ProjectionSpec( |
| 200 | 7168, |
| 201 | 4224, |
| 202 | cute_configs=((1, _cute(1, 96, 4, 2, 4)),), |
| 203 | name="dense_down_proj", |
| 204 | ), |
| 205 | (7168, 8448): ProjectionSpec( |
| 206 | 7168, |
| 207 | 8448, |
| 208 | cute_configs=( |
| 209 | (1, _cute(1, 32, 4, 4)), |
| 210 | (2, _cute(2, 96, 4, 1)), |
| 211 | (3, _cute(3, 96, 4, 1)), |
| 212 | ), |
| 213 | name="dense_down_proj", |
| 214 | ), |
| 215 | (8448, 7168): ProjectionSpec( |
| 216 | 8448, |
| 217 | 7168, |
| 218 | cute_configs=( |
| 219 | (1, _cute(1, 224, 3, 4)), |
| 220 | (2, _cute(2, 32, 4, 4)), |
| 221 | ), |
| 222 | name="dense_gate_up_proj", |
| 223 | ), |
| 224 | (16896, 7168): ProjectionSpec( |
| 225 | 16896, |
| 226 | 7168, |
| 227 | cute_configs=( |
| 228 | (1, _cute(1, 224, 6, 4)), |
| 229 | (2, _cute(2, 32, 4, 4)), |
| 230 | ), |
| 231 | name="dense_gate_up_proj", |
| 232 | ), |
| 233 | (7168, 14336): ProjectionSpec( |
| 234 | 7168, |
| 235 | 14336, |
| 236 | cute_configs=( |
| 237 | (1, SkinnyGemmConfig(1, 256, 2, vector_width=4, static_k=14336)), |
| 238 | (2, _cute(2, 224, 4, 2)), |
| 239 | ), |
| 240 | ), |
| 241 | (20480, 7168): ProjectionSpec( |
| 242 | 20480, |
| 243 | 7168, |
| 244 | cute_configs=( |
| 245 | (1, _cute(1, 224, 4, 2)), |
| 246 | (2, _cute(2, 64, 4, 2)), |
| 247 | (3, _cute(3, 64, 2, 2)), |
| 248 | (4, _cute(4, 64, 4, 1)), |
| 249 | ), |
| 250 | name="lm_head", |
| 251 | ), |
| 252 | (40960, 7168): ProjectionSpec( |
| 253 | 40960, |
| 254 | 7168, |
| 255 | cute_configs=( |
| 256 | (1, _cute(1, 128, 4, 2)), |
| 257 | (2, _cute(2, 64, 4, 2)), |
| 258 | (3, _cute(3, 64, 2, 2)), |
| 259 | (4, _cute(4, 64, 4, 1)), |
| 260 | ), |
| 261 | name="lm_head", |
| 262 | ), |
| 263 | # TP16. Measured on B300 over M=1..16 with the same >=5% threshold as the |
| 264 | # entries above. The replicated projections (2112x7168, 3584x7168, |
| 265 | # 7168x3584) keep their shapes at TP16 and reuse the entries above, and |
| 266 | # o_proj lands on 7168x768, which shared_down_proj already covers. |
| 267 | (3216, 7168): ProjectionSpec( |
| 268 | 3216, |
| 269 | 7168, |
| 270 | # Both gaps in this range are measured, not oversights: dsv3 is only |
| 271 | # 4% ahead at M6..M8, and at M16 cuBLAS switches to a faster kernel |
| 272 | # (11.42us vs dsv3's 11.83us) after trailing it by 6-8% at M9..M15. |
| 273 | frozenset(range(9, 16)), |
| 274 | cute_configs=( |
| 275 | (1, _cute(1, 224, 3, 4)), |
| 276 | (2, _cute(2, 128, 4, 2)), |
| 277 | (3, _cute(3, 128, 2, 1)), |
| 278 | (4, _cute(4, 64, 2, 2)), |
| 279 | (5, _cute(5, 128, 3, 1)), |
| 280 | ), |
| 281 | name="in_proj_qkvgfab", |
| 282 | ), |
| 283 | (768, 7168): ProjectionSpec( |
| 284 | 768, |
| 285 | 7168, |
| 286 | frozenset(range(5, 17)), |
| 287 | cute_configs=( |
| 288 | (1, _cute(1, 224, 2, 4)), |
| 289 | (2, _cute(2, 224, 2, 2)), |
| 290 | (3, _cute(3, 224, 2, 2)), |
| 291 | (4, _cute(4, 224, 2, 2)), |
| 292 | ), |
| 293 | name="mla_g_proj/shared_gate_up_proj", |
| 294 | ), |
| 295 | (1152, 1536): ProjectionSpec( |
| 296 | 1152, |
| 297 | 1536, |
| 298 | frozenset(range(2, 17)), |
| 299 | ((1, _cute(1, 192, 3, 4)),), |
| 300 | name="q_b_proj", |
| 301 | ), |
| 302 | (768, 128): ProjectionSpec(768, 128, _M1_TO_16, name="f_b_proj"), |
| 303 | # dsv3 drops under 5% from M9 on for this shape. |
| 304 | (7168, 384): ProjectionSpec( |
| 305 | 7168, 384, frozenset(range(1, 9)), name="shared_down_proj" |
| 306 | ), |
| 307 | (4224, 7168): ProjectionSpec( |
| 308 | 4224, |
| 309 | 7168, |
| 310 | frozenset(range(4, 9)), |
| 311 | cute_configs=( |
| 312 | (1, _cute(1, 224, 3, 4)), |
| 313 | (2, _cute(2, 128, 2, 1)), |
| 314 | (3, _cute(3, 64, 2, 2)), |
| 315 | ), |
| 316 | name="dense_gate_up_proj", |
| 317 | ), |
| 318 | (10240, 7168): ProjectionSpec( |
| 319 | 10240, |
| 320 | 7168, |
| 321 | cute_configs=( |
| 322 | (1, _cute(1, 224, 4, 2)), |
| 323 | (2, _cute(2, 32, 2, 4)), |
| 324 | (3, _cute(3, 64, 4, 1)), |
| 325 | (4, _cute(4, 64, 4, 1)), |
| 326 | ), |
| 327 | name="lm_head", |
| 328 | ), |
| 329 | # 7168x2112 (TP16 dense down_proj) has no entry on purpose: K=2112 divides |
| 330 | # none of the fused-A tile_k values, and the CuTe kernel is left with |
| 331 | # vector_width=2, which measured slower than cuBLAS. |
| 332 | } |
| 333 | |
| 334 | # Keyed by local (N, K), B200 (SM100) entries. Measured on B200 at TP8/TP16 |
| 335 | # shapes over M=1..16 with the same >=5%-over-cuBLAS threshold as the SM103 |
| 336 | # table. Per-(shape, M) entries take the best of the SM103-tuned config and a |
| 337 | # fresh SM100 measurement: several SM103 configs (7168x8448, 20480x7168, |
| 338 | # 40960x7168, 10240x7168, 7168x3072 M2, 7168x1536) beat the heuristic config on |
| 339 | # B200 outright, while a few (6288x7168 M2..4, 3072x7168 M3..5) lose to cuBLAS |
| 340 | # on B200 and are intentionally not inherited. The 7168x3584 residual configs |
| 341 | # are intentionally absent: the B200 residual path was not measured. |
| 342 | KIMI_K3_PROJECTIONS_SM100: dict[tuple[int, int], ProjectionSpec] = { |
| 343 | (1536, 128): ProjectionSpec(1536, 128, frozenset({1, 16}), name="f_b_proj"), |
| 344 | (3072, 128): ProjectionSpec(3072, 128, frozenset({8}), name="f_b_proj"), |
| 345 | # Same mla_g_proj aux-stream/PDL capture caveat as the SM103 entry applies; |
| 346 | # only M4/M8 measured a dsv3 win, so this is the narrow set either way. |
| 347 | (1536, 7168): ProjectionSpec( |
| 348 | 1536, |
| 349 | 7168, |
| 350 | frozenset({4, 8}), |
| 351 | ( |
| 352 | (1, _cute(1, 224, 3, 2)), |
| 353 | (2, _cute(2, 128, 2, 2)), |
| 354 | ), |
| 355 | name="shared_gate_up_proj/mla_g_proj", |
| 356 | ), |
| 357 | (3072, 7168): ProjectionSpec( |
| 358 | 3072, |
| 359 | 7168, |
| 360 | cute_configs=( |
| 361 | (1, _cute(1, 224, 3, 4)), |
| 362 | (2, _cute(2, 128, 2, 1)), |
| 363 | ), |
| 364 | name="shared_gate_up_proj", |
| 365 | ), |
| 366 | (2112, 7168): ProjectionSpec( |
| 367 | 2112, |
| 368 | 7168, |
| 369 | frozenset({4, 16}), |
| 370 | ( |
| 371 | (1, _cute(1, 224, 3, 2)), |
| 372 | (2, _cute(2, 128, 2, 2)), |
| 373 | ), |
| 374 | name="fused_qkv_a_proj", |
| 375 | ), |
| 376 | (2304, 1536): ProjectionSpec(2304, 1536, _M1_TO_16, name="q_b_proj"), |
| 377 | (4608, 1536): ProjectionSpec(4608, 1536, frozenset({1, 2, 4}), name="q_b_proj"), |
| 378 | # M2..4 of the SM103 configs lose to cuBLAS on B200; only M1 carries over. |
| 379 | (6288, 7168): ProjectionSpec( |
| 380 | 6288, |
| 381 | 7168, |
| 382 | cute_configs=((1, _cute(1, 224, 3, 4)),), |
| 383 | name="in_proj_qkvgfab", |
| 384 | ), |
| 385 | (12448, 7168): ProjectionSpec( |
| 386 | 12448, |
| 387 | 7168, |
| 388 | cute_configs=( |
| 389 | (1, _cute(1, 224, 4, 2)), |
| 390 | (2, _cute(2, 64, 4, 2)), |
| 391 | (3, _cute(3, 64, 2, 2)), |
| 392 | (4, _cute(4, 128, 2, 1)), |
| 393 | ), |
| 394 | name="in_proj_qkvgfab", |
| 395 | ), |
| 396 | (7168, 768): ProjectionSpec(7168, 768, frozenset({1}), name="shared_down_proj"), |
| 397 | # SM103-tuned config wins on B200 (+25%) where the heuristic config tied. |
| 398 | (7168, 1536): ProjectionSpec( |
| 399 | 7168, |
| 400 | 1536, |
| 401 | cute_configs=((1, _cute(1, 96, 4, 2)),), |
| 402 | name="o_proj", |
| 403 | ), |
| 404 | (7168, 3072): ProjectionSpec( |
| 405 | 7168, |
| 406 | 3072, |
| 407 | cute_configs=( |
| 408 | (1, _cute(1, 96, 2, 4)), |
| 409 | (2, _cute(2, 32, 4, 4)), |
| 410 | ), |
| 411 | name="o_proj", |
| 412 | ), |
| 413 | (7168, 3584): ProjectionSpec( |
| 414 | 7168, |
| 415 | 3584, |
| 416 | cute_configs=( |
| 417 | (1, _cute(1, 64, 4, 2)), |
| 418 | (2, _cute(2, 64, 4, 2)), |
| 419 | ), |
| 420 | name="routed_expert_up_proj", |
| 421 | ), |
| 422 | (7168, 8448): ProjectionSpec( |
| 423 | 7168, |
| 424 | 8448, |
| 425 | cute_configs=( |
| 426 | (1, _cute(1, 32, 4, 4)), |
| 427 | (2, _cute(2, 96, 4, 1)), |
| 428 | (3, _cute(3, 96, 4, 1)), |
| 429 | ), |
| 430 | name="dense_down_proj", |
| 431 | ), |
| 432 | (8448, 7168): ProjectionSpec( |
| 433 | 8448, |
| 434 | 7168, |
| 435 | cute_configs=( |
| 436 | (1, _cute(1, 224, 3, 4)), |
| 437 | (2, _cute(2, 32, 4, 4)), |
| 438 | ), |
| 439 | name="dense_gate_up_proj", |
| 440 | ), |
| 441 | (16896, 7168): ProjectionSpec( |
| 442 | 16896, |
| 443 | 7168, |
| 444 | cute_configs=((1, _cute(1, 224, 3, 4)),), |
| 445 | name="dense_gate_up_proj", |
| 446 | ), |
| 447 | (20480, 7168): ProjectionSpec( |
| 448 | 20480, |
| 449 | 7168, |
| 450 | cute_configs=( |
| 451 | (1, _cute(1, 224, 4, 2)), |
| 452 | (2, _cute(2, 64, 4, 2)), |
| 453 | (3, _cute(3, 64, 2, 2)), |
| 454 | (4, _cute(4, 64, 4, 1)), |
| 455 | ), |
| 456 | name="lm_head", |
| 457 | ), |
| 458 | (40960, 7168): ProjectionSpec( |
| 459 | 40960, |
| 460 | 7168, |
| 461 | cute_configs=( |
| 462 | (1, _cute(1, 128, 4, 2)), |
| 463 | (2, _cute(2, 64, 4, 2)), |
| 464 | (3, _cute(3, 64, 4, 1)), |
| 465 | (4, _cute(4, 64, 4, 1)), |
| 466 | ), |
| 467 | name="lm_head", |
| 468 | ), |
| 469 | (10240, 7168): ProjectionSpec( |
| 470 | 10240, |
| 471 | 7168, |
| 472 | cute_configs=( |
| 473 | (1, _cute(1, 224, 4, 2)), |
| 474 | (2, _cute(2, 32, 2, 4)), |
| 475 | (3, _cute(3, 64, 4, 1)), |
| 476 | (4, _cute(4, 64, 4, 1)), |
| 477 | ), |
| 478 | name="lm_head", |
| 479 | ), |
| 480 | # Latent-MoE replicated projections. dsv3 measured M2..8 wins on B300 for |
| 481 | # 3584x7168 but loses to cuBLAS there on B200; only CuTe wins land here. |
| 482 | (3584, 7168): ProjectionSpec( |
| 483 | 3584, |
| 484 | 7168, |
| 485 | cute_configs=( |
| 486 | (1, _cute(1, 224, 2, 4)), |
| 487 | (2, _cute(2, 128, 2, 1)), |
| 488 | (3, _cute(3, 128, 2, 1)), |
| 489 | ), |
| 490 | name="routed_expert_down_proj", |
| 491 | ), |
| 492 | # TP16 shapes, newly measured on B200 (they only appear in TP16 deployments; |
| 493 | # on TP8 these table entries simply never match). |
| 494 | (768, 7168): ProjectionSpec( |
| 495 | 768, |
| 496 | 7168, |
| 497 | cute_configs=( |
| 498 | (1, _cute(1, 224, 2, 4)), |
| 499 | (2, _cute(2, 224, 2, 2)), |
| 500 | (3, _cute(3, 224, 2, 2)), |
| 501 | (4, _cute(4, 224, 2, 2)), |
| 502 | ), |
| 503 | name="mla_g_proj/shared_gate_up_proj", |
| 504 | ), |
| 505 | (1152, 1536): ProjectionSpec( |
| 506 | 1152, |
| 507 | 1536, |
| 508 | cute_configs=((1, _cute(1, 192, 3, 4)),), |
| 509 | name="q_b_proj", |
| 510 | ), |
| 511 | (3216, 7168): ProjectionSpec( |
| 512 | 3216, |
| 513 | 7168, |
| 514 | cute_configs=( |
| 515 | (1, _cute(1, 224, 3, 4)), |
| 516 | (2, _cute(2, 128, 4, 2)), |
| 517 | ), |
| 518 | name="in_proj_qkvgfab", |
| 519 | ), |
| 520 | (4224, 7168): ProjectionSpec( |
| 521 | 4224, |
| 522 | 7168, |
| 523 | cute_configs=( |
| 524 | (1, _cute(1, 224, 3, 4)), |
| 525 | (2, _cute(2, 128, 2, 1)), |
| 526 | (3, _cute(3, 64, 2, 2)), |
| 527 | ), |
| 528 | name="dense_gate_up_proj", |
| 529 | ), |
| 530 | } |
| 531 | |
| 532 | _SM90CuteConfig = tuple[int, int, int, int] |
| 533 | |
| 534 | |
| 535 | def _sm90_spec( |
| 536 | n: int, |
| 537 | k: int, |
| 538 | dsv3_tokens: frozenset[int] = frozenset(), |
| 539 | configs: tuple[_SM90CuteConfig, ...] = (), |
| 540 | ) -> ProjectionSpec: |
| 541 | cute_configs = tuple( |
| 542 | (m, _cute(m, *config)) for m, config in enumerate(configs, start=1) |
| 543 | ) |
| 544 | return ProjectionSpec(n, k, dsv3_tokens, cute_configs) |
| 545 | |
| 546 | |
| 547 | KIMI_K3_PROJECTIONS_SM90: dict[tuple[int, int], ProjectionSpec] = { |
| 548 | (1536, 128): _sm90_spec(1536, 128, frozenset(range(1, 9))), |
| 549 | (3072, 128): _sm90_spec(3072, 128, frozenset({1, 2, 5, 6, 7, 8, 9})), |
| 550 | (1536, 7168): _sm90_spec( |
| 551 | 1536, |
| 552 | 7168, |
| 553 | frozenset(range(7, 17)), |
| 554 | ( |
| 555 | (224, 2, 4, 8), |
| 556 | (128, 3, 2, 8), |
| 557 | (128, 2, 1, 8), |
| 558 | (128, 2, 1, 8), |
| 559 | (128, 3, 1, 8), |
| 560 | (128, 3, 1, 8), |
| 561 | ), |
| 562 | ), |
| 563 | (3072, 7168): _sm90_spec( |
| 564 | 3072, |
| 565 | 7168, |
| 566 | configs=((224, 2, 4, 8), (128, 2, 2, 8), (64, 4, 1, 8), (128, 6, 1, 8)), |
| 567 | ), |
| 568 | (2112, 7168): _sm90_spec( |
| 569 | 2112, |
| 570 | 7168, |
| 571 | frozenset(range(5, 17)), |
| 572 | ((224, 2, 4, 8), (128, 4, 2, 8), (128, 2, 1, 8), (128, 2, 1, 8)), |
| 573 | ), |
| 574 | (2304, 1536): _sm90_spec( |
| 575 | 2304, |
| 576 | 1536, |
| 577 | frozenset(range(3, 9)), |
| 578 | ((96, 4, 2, 8), (96, 4, 1, 8)), |
| 579 | ), |
| 580 | (4608, 1536): _sm90_spec( |
| 581 | 4608, |
| 582 | 1536, |
| 583 | configs=((96, 4, 2, 4), (96, 4, 1, 8), (64, 4, 1, 8)), |
| 584 | ), |
| 585 | (3584, 7168): _sm90_spec( |
| 586 | 3584, |
| 587 | 7168, |
| 588 | configs=((224, 2, 4, 8), (128, 4, 2, 8), (128, 2, 1, 8)), |
| 589 | ), |
| 590 | (6288, 7168): _sm90_spec( |
| 591 | 6288, |
| 592 | 7168, |
| 593 | configs=((224, 2, 4, 8), (128, 4, 2, 8), (64, 4, 1, 8)), |
| 594 | ), |
| 595 | (12448, 7168): _sm90_spec( |
| 596 | 12448, |
| 597 | 7168, |
| 598 | configs=((224, 2, 4, 8), (224, 4, 2, 8), (128, 2, 1, 8)), |
| 599 | ), |
| 600 | (7168, 768): _sm90_spec(7168, 768, configs=((96, 4, 2, 4), (96, 4, 1, 8))), |
| 601 | (7168, 1536): _sm90_spec(7168, 1536, configs=((96, 4, 2, 8), (96, 4, 1, 8))), |
| 602 | (7168, 3072): _sm90_spec( |
| 603 | 7168, |
| 604 | 3072, |
| 605 | configs=((96, 2, 4, 8), (64, 4, 2, 8), (64, 2, 2, 8)), |
| 606 | ), |
| 607 | (7168, 3584): _sm90_spec( |
| 608 | 7168, |
| 609 | 3584, |
| 610 | configs=((224, 4, 2, 8), (64, 4, 2, 8), (64, 4, 2, 8)), |
| 611 | ), |
| 612 | (7168, 4224): _sm90_spec(7168, 4224, configs=((96, 4, 2, 4), (32, 4, 2, 4))), |
| 613 | (7168, 8448): _sm90_spec(7168, 8448, configs=((96, 2, 4, 8), (32, 4, 2, 8))), |
| 614 | (7168, 14336): _sm90_spec(7168, 14336, configs=((224, 2, 4, 8), (128, 2, 2, 8))), |
| 615 | (20480, 7168): _sm90_spec( |
| 616 | 20480, |
| 617 | 7168, |
| 618 | configs=((224, 4, 2, 8), (224, 4, 2, 8), (128, 2, 1, 8)), |
| 619 | ), |
| 620 | } |
| 621 | |
| 622 | |
| 623 | def _backend_for( |
| 624 | spec: ProjectionSpec, num_tokens: int, has_residual: bool |
| 625 | ) -> Backend | None: |
| 626 | if has_residual: |
| 627 | return "cute" if spec.residual_config(num_tokens) is not None else None |
| 628 | if spec.cute_config(num_tokens) is not None: |
| 629 | return "cute" |
| 630 | if num_tokens in spec.dsv3_tokens: |
| 631 | return "dsv3_fused_a" |
| 632 | return None |
| 633 | |
| 634 | |
| 635 | def select_kimi_k3_backend( |
| 636 | num_tokens: int, |
| 637 | n: int, |
| 638 | k: int, |
| 639 | *, |
| 640 | has_residual: bool = False, |
| 641 | ) -> Backend | None: |
| 642 | """Backend for a local ``(N, K)`` at ``num_tokens``, or None to fall back.""" |
| 643 | spec = KIMI_K3_PROJECTIONS.get((n, k)) |
| 644 | return _backend_for(spec, num_tokens, has_residual) if spec is not None else None |
| 645 | |
| 646 | |
| 647 | def _build_plan(spec: ProjectionSpec) -> dict[int, ResolvedCall]: |
| 648 | plan: dict[int, ResolvedCall] = {} |
| 649 | for num_tokens in range(1, 17): |
| 650 | backend = _backend_for(spec, num_tokens, has_residual=False) |
| 651 | if backend == "cute": |
| 652 | plan[num_tokens] = ("cute", spec.cute_config(num_tokens)) |
| 653 | elif backend == "dsv3_fused_a": |
| 654 | plan[num_tokens] = ("dsv3_fused_a", None) |
| 655 | return plan |
| 656 | |
| 657 | |
| 658 | def _build_residual_plan(spec: ProjectionSpec) -> dict[int, SkinnyGemmConfig]: |
| 659 | return {num_tokens: config for num_tokens, config in spec.residual_configs} |
| 660 | |
| 661 | |
| 662 | def _is_sm103() -> bool: |
| 663 | return current_platform.is_device_capability((10, 3)) |
| 664 | |
| 665 | |
| 666 | def _is_sm107() -> bool: |
| 667 | return current_platform.is_device_capability((10, 7)) |
| 668 | |
| 669 | |
| 670 | def _low_latency_table() -> dict[tuple[int, int], ProjectionSpec] | None: |
| 671 | """Measured dispatch table for the current device, or None if unsupported.""" |
| 672 | if _is_sm103() or _is_sm107(): |
| 673 | # SM107 reuses the SM103 table; see the module docstring. |
| 674 | return KIMI_K3_PROJECTIONS |
| 675 | if current_platform.is_device_capability((10, 0)): |
| 676 | return KIMI_K3_PROJECTIONS_SM100 |
| 677 | if current_platform.is_device_capability((9, 0)): |
| 678 | return KIMI_K3_PROJECTIONS_SM90 |
| 679 | return None |
| 680 | |
| 681 | |
| 682 | def _is_packed_row_major(tensor: torch.Tensor) -> bool: |
| 683 | return tensor.dim() == 2 and tensor.stride() == (tensor.shape[1], 1) |
| 684 | |
| 685 | |
| 686 | def _runtime_ok(x: torch.Tensor, weight: torch.Tensor) -> bool: |
| 687 | return ( |
| 688 | x.dim() == 2 |
| 689 | and x.stride(1) == 1 |
| 690 | and (x.shape[0] == 1 or x.stride(0) % 8 == 0) |
| 691 | and _is_packed_row_major(weight) |
| 692 | and x.dtype == torch.bfloat16 |
| 693 | and weight.dtype == torch.bfloat16 |
| 694 | and x.is_cuda |
| 695 | and weight.is_cuda |
| 696 | and x.device == weight.device |
| 697 | and x.shape[1] == weight.shape[1] |
| 698 | ) |
| 699 | |
| 700 | |
| 701 | def _residual_ok(x: torch.Tensor, weight: torch.Tensor, residual: torch.Tensor) -> bool: |
| 702 | return ( |
| 703 | _is_packed_row_major(x) |
| 704 | and residual.dim() == 2 |
| 705 | and residual.dtype == torch.bfloat16 |
| 706 | and residual.is_cuda |
| 707 | and residual.device == x.device |
| 708 | and residual.is_contiguous() |
| 709 | and residual.shape == (x.shape[0], weight.shape[0]) |
| 710 | ) |
| 711 | |
| 712 | |
| 713 | def _run_plan( |
| 714 | plan: dict[int, ResolvedCall], x: torch.Tensor, weight: torch.Tensor |
| 715 | ) -> torch.Tensor | None: |
| 716 | entry = plan.get(x.shape[0]) |
| 717 | if entry is None: |
| 718 | return None |
| 719 | backend, config = entry |
| 720 | if backend == "cute": |
| 721 | if x.shape[0] != 1 and not _is_packed_row_major(x): |
| 722 | return None |
| 723 | if not shape_dynamic_skinny_gemm.is_available(): |
| 724 | return None |
| 725 | return shape_dynamic_skinny_gemm(x, weight, config, None) |
| 726 | if not hasattr(torch.ops._C, "dsv3_fused_a_gemm"): |
| 727 | return None |
| 728 | if x.shape[0] != 1 and not _is_packed_row_major(x): |
| 729 | return None |
| 730 | output = torch.empty((x.shape[0], weight.shape[0]), dtype=x.dtype, device=x.device) |
| 731 | ops.dsv3_fused_a_gemm(output, x, weight.t(), enable_pdl=True) |
| 732 | return output |
| 733 | |
| 734 | |
| 735 | def run_kda_projection_overlap( |
| 736 | hidden_states: torch.Tensor, |
| 737 | packed_weight: torch.Tensor, |
| 738 | f_b_weight: torch.Tensor, |
| 739 | aux_stream: torch.cuda.Stream, |
| 740 | events: tuple[torch.cuda.Event, torch.cuda.Event], |
| 741 | ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: |
| 742 | """Run the TP8 KDA decode projection branches concurrently. |
| 743 | |
| 744 | Args: |
| 745 | hidden_states: Packed BF16 input with shape ``[M, 7168]``. |
| 746 | packed_weight: Existing Q/K/V/G/F_A/beta/pad weight with shape |
| 747 | ``[6288, 7168]``. |
| 748 | f_b_weight: F_B weight with shape ``[1536, 128]``. |
| 749 | aux_stream: Stream for the F_A/beta then F_B branch. |
| 750 | events: Start and completion events for the stream fork and join. |
| 751 | |
| 752 | Returns: |
| 753 | The Q/K/V/G projection, F_B output, and beta projection. |
| 754 | |
| 755 | """ |
| 756 | num_tokens = hidden_states.shape[0] |
| 757 | qkvg_weight = packed_weight[:_KDA_QKVG_SIZE] |
| 758 | |
| 759 | def run_qkvg() -> torch.Tensor: |
| 760 | if config := KDA_QKVG_CONFIGS.get(num_tokens): |
| 761 | return shape_dynamic_skinny_gemm( |
| 762 | hidden_states, |
| 763 | qkvg_weight, |
| 764 | config, |
| 765 | None, |
| 766 | ) |
| 767 | if num_tokens <= KDA_PROJECTION_OVERLAP_MAX_TOKENS and ( |
| 768 | backend := _kda_qkvg_flashinfer_backend() |
| 769 | ): |
| 770 | from flashinfer.gemm import mm_bf16 |
| 771 | |
| 772 | return mm_bf16( |
| 773 | hidden_states, |
| 774 | qkvg_weight.t(), |
| 775 | pdl=True, |
| 776 | backend=backend, |
| 777 | ) |
| 778 | return torch.mm(hidden_states, qkvg_weight.t()) |
| 779 | |
| 780 | def run_fab_fb() -> tuple[torch.Tensor, torch.Tensor]: |
| 781 | if num_tokens == 1: |
| 782 | projected_fab = shape_dynamic_skinny_gemm( |
| 783 | hidden_states, |
| 784 | packed_weight[_KDA_QKVG_SIZE : _KDA_QKVG_SIZE + _KDA_FAB_SIZE], |
| 785 | KDA_M1_FAB_CONFIG, |
| 786 | None, |
| 787 | ) |
| 788 | f_a, beta = projected_fab.split([128, 12], dim=-1) |
| 789 | f_a = f_a.as_strided((1, 128), (128, 1)) |
| 790 | g1 = torch.empty( |
| 791 | (1, 1536), dtype=hidden_states.dtype, device=hidden_states.device |
| 792 | ) |
| 793 | ops.dsv3_fused_a_gemm(g1, f_a, f_b_weight.t(), enable_pdl=True) |
| 794 | else: |
| 795 | from vllm.models.kimi_k3.nvidia.ops.cute_dsl.kda_skinny_gemm import ( |
| 796 | kda_skinny_gemm, |
| 797 | ) |
| 798 | |
| 799 | if num_tokens <= KDA_SKINNY_N_MAX_TOKENS: |
| 800 | projected_fab = kda_skinny_gemm.run_n( |
| 801 | hidden_states, |
| 802 | packed_weight[_KDA_QKVG_SIZE:], |
| 803 | ) |
| 804 | else: |
| 805 | projected_fab = torch.mm( |
| 806 | hidden_states, |
| 807 | packed_weight[_KDA_QKVG_SIZE:].t(), |
| 808 | ) |
| 809 | beta = projected_fab[:, 128:140] |
| 810 | if num_tokens <= KDA_SKINNY_K_MAX_TOKENS: |
| 811 | g1 = kda_skinny_gemm.run_k(projected_fab, f_b_weight) |
| 812 | else: |
| 813 | g1 = torch.mm(projected_fab[:, :128], f_b_weight.t()) |
| 814 | return g1, beta |
| 815 | |
| 816 | projected_qkvg, (g1, beta) = maybe_execute_in_parallel( |
| 817 | run_qkvg, |
| 818 | run_fab_fb, |
| 819 | events[0], |
| 820 | events[1], |
| 821 | aux_stream, |
| 822 | ) |
| 823 | return projected_qkvg, g1, beta |
| 824 | |
| 825 | |
| 826 | def autotune_kda_qkvg(model: nn.Module) -> None: |
| 827 | """Autotune the supported QKVG GEMM before CUDA graph capture.""" |
| 828 | from vllm.models.kimi_k3.nvidia.kda import KimiK3DeltaAttention |
| 829 | |
| 830 | backend = _kda_qkvg_flashinfer_backend() |
| 831 | children: list[KimiK3DeltaAttention] = [] |
| 832 | weights_by_shape: dict[ |
| 833 | tuple[int, int, int, torch.dtype, torch.device], torch.Tensor |
| 834 | ] = {} |
| 835 | for child in model.modules(): |
| 836 | if not isinstance(child, KimiK3DeltaAttention): |
| 837 | continue |
| 838 | if child._projection_overlap_max_tokens <= 0: |
| 839 | continue |
| 840 | children.append(child) |
| 841 | if backend is not None: |
| 842 | qkvg_weight = child.in_proj_qkvgfab.weight[:_KDA_QKVG_SIZE] |
| 843 | shape = ( |
| 844 | KDA_PROJECTION_OVERLAP_MAX_TOKENS, |
| 845 | qkvg_weight.shape[0], |
| 846 | qkvg_weight.shape[1], |
| 847 | qkvg_weight.dtype, |
| 848 | qkvg_weight.device, |
| 849 | ) |
| 850 | weights_by_shape.setdefault(shape, qkvg_weight) |
| 851 | |
| 852 | if backend is not None: |
| 853 | from flashinfer.gemm import mm_bf16 |
| 854 | |
| 855 | for shape, qkvg_weight in weights_by_shape.items(): |
| 856 | num_tokens = shape[0] |
| 857 | hidden_states = torch.empty( |
| 858 | (num_tokens, qkvg_weight.shape[1]), |
| 859 | dtype=qkvg_weight.dtype, |
| 860 | device=qkvg_weight.device, |
| 861 | ) |
| 862 | mm_bf16( |
| 863 | hidden_states, |
| 864 | qkvg_weight.t(), |
| 865 | pdl=True, |
| 866 | backend=backend, |
| 867 | ) |
| 868 | for child in children: |
| 869 | child._projection_overlap_max_tokens = KDA_PROJECTION_OVERLAP_MAX_TOKENS |
| 870 | |
| 871 | |
| 872 | def _run_residual_plan( |
| 873 | residual_plan: dict[int, SkinnyGemmConfig], |
| 874 | x: torch.Tensor, |
| 875 | weight: torch.Tensor, |
| 876 | residual: torch.Tensor, |
| 877 | ) -> torch.Tensor | None: |
| 878 | config = residual_plan.get(x.shape[0]) |
| 879 | if config is None or not shape_dynamic_skinny_gemm.is_available(): |
| 880 | return None |
| 881 | return shape_dynamic_skinny_gemm(x, weight, config, residual) |
| 882 | |
| 883 | |
| 884 | def try_low_latency_gemm( |
| 885 | x: torch.Tensor, |
| 886 | weight: torch.Tensor, |
| 887 | residual: torch.Tensor | None = None, |
| 888 | ) -> torch.Tensor | None: |
| 889 | """Run the shape-selected low-latency kernel, or None to fall back. |
| 890 | |
| 891 | Resolves the plan from the shape table on each call; production installs a |
| 892 | precomputed plan (see :func:`enable_kimi_k3_low_latency_gemm`) and does not |
| 893 | use this path. |
| 894 | """ |
| 895 | if envs.VLLM_BATCH_INVARIANT or not _runtime_ok(x, weight): |
| 896 | return None |
| 897 | table = _low_latency_table() |
| 898 | if table is None: |
| 899 | return None |
| 900 | spec = table.get((weight.shape[0], weight.shape[1])) |
| 901 | if spec is None: |
| 902 | return None |
| 903 | if residual is None: |
| 904 | return _run_plan(_build_plan(spec), x, weight) |
| 905 | if not _residual_ok(x, weight, residual): |
| 906 | return None |
| 907 | return _run_residual_plan(_build_residual_plan(spec), x, weight, residual) |
| 908 | |
| 909 | |
| 910 | class _KimiK3LowLatencyApply: |
| 911 | """Mixin: try the precomputed plan, else defer to the base method.""" |
| 912 | |
| 913 | def __init__(self, plan: dict[int, ResolvedCall]) -> None: |
| 914 | super().__init__() |
| 915 | self._plan = plan |
| 916 | |
| 917 | def apply( |
| 918 | self, |
| 919 | layer: nn.Module, |
| 920 | x: torch.Tensor, |
| 921 | bias: torch.Tensor | None = None, |
| 922 | ) -> torch.Tensor: |
| 923 | if ( |
| 924 | bias is None |
| 925 | and not envs.VLLM_BATCH_INVARIANT |
| 926 | and _runtime_ok(x, layer.weight) |
| 927 | ): |
| 928 | output = _run_plan(self._plan, x, layer.weight) |
| 929 | if output is not None: |
| 930 | return output |
| 931 | return super().apply(layer, x, bias) # type: ignore[misc] |
| 932 | |
| 933 | |
| 934 | class KimiK3LowLatencyLinearMethod(_KimiK3LowLatencyApply, UnquantizedLinearMethod): |
| 935 | def __init__( |
| 936 | self, |
| 937 | plan: dict[int, ResolvedCall], |
| 938 | residual_plan: dict[int, SkinnyGemmConfig], |
| 939 | ) -> None: |
| 940 | super().__init__(plan) |
| 941 | self._residual_plan = residual_plan |
| 942 | |
| 943 | def apply_with_residual( |
| 944 | self, |
| 945 | layer: nn.Module, |
| 946 | x: torch.Tensor, |
| 947 | residual: torch.Tensor, |
| 948 | ) -> torch.Tensor: |
| 949 | if ( |
| 950 | not envs.VLLM_BATCH_INVARIANT |
| 951 | and _runtime_ok(x, layer.weight) |
| 952 | and _residual_ok(x, layer.weight, residual) |
| 953 | ): |
| 954 | output = _run_residual_plan(self._residual_plan, x, layer.weight, residual) |
| 955 | if output is not None: |
| 956 | return output |
| 957 | return torch.addmm(residual, x, layer.weight.t()) |
| 958 | |
| 959 | |
| 960 | class KimiK3LowLatencyEmbeddingMethod( |
| 961 | _KimiK3LowLatencyApply, UnquantizedEmbeddingMethod |
| 962 | ): |
| 963 | pass |
| 964 | |
| 965 | |
| 966 | def _enable_kda_projection_overlap(module: nn.Module) -> bool: |
| 967 | from vllm.models.kimi_k3.nvidia.kda import KimiK3DeltaAttention |
| 968 | |
| 969 | if envs.VLLM_BATCH_INVARIANT: |
| 970 | return False |
| 971 | |
| 972 | enabled = False |
| 973 | for child in module.modules(): |
| 974 | if not isinstance(child, KimiK3DeltaAttention): |
| 975 | continue |
| 976 | if child.tp_size != _KDA_TP_SIZE: |
| 977 | continue |
| 978 | packed_weight = child.in_proj_qkvgfab.weight |
| 979 | f_b_weight = child.f_b_proj.weight |
| 980 | if ( |
| 981 | type(child.in_proj_qkvgfab.quant_method) is not KimiK3LowLatencyLinearMethod |
| 982 | or type(child.f_b_proj.quant_method) is not KimiK3LowLatencyLinearMethod |
| 983 | or child._projection_aux_stream is None |
| 984 | or child._projection_events is None |
| 985 | or packed_weight.shape != (_KDA_PACKED_SIZE, 7168) |
| 986 | or f_b_weight.shape != (1536, 128) |
| 987 | or packed_weight.dtype != torch.bfloat16 |
| 988 | or f_b_weight.dtype != torch.bfloat16 |
| 989 | or not packed_weight.is_cuda |
| 990 | or not f_b_weight.is_cuda |
| 991 | or not packed_weight.is_contiguous() |
| 992 | or not f_b_weight.is_contiguous() |
| 993 | ): |
| 994 | continue |
| 995 | child._projection_overlap_max_tokens = max(KDA_QKVG_CONFIGS) |
| 996 | enabled = True |
| 997 | return enabled |
| 998 | |
| 999 | |
| 1000 | def enable_kimi_k3_low_latency_gemm( |
| 1001 | module: nn.Module, |
| 1002 | dtype: torch.dtype, |
| 1003 | ) -> None: |
| 1004 | """Install shape-selected low-latency GEMMs and register CuTe warmups. |
| 1005 | |
| 1006 | Modules are matched purely by type, an exactly-unquantized method, and a |
| 1007 | local ``(N, K)`` present in the current device's measured table |
| 1008 | (:data:`KIMI_K3_PROJECTIONS` on SM103, :data:`KIMI_K3_PROJECTIONS_SM100` |
| 1009 | on SM100, :data:`KIMI_K3_PROJECTIONS_SM90` on SM90). |
| 1010 | """ |
| 1011 | if dtype != torch.bfloat16: |
| 1012 | return |
| 1013 | table = _low_latency_table() |
| 1014 | if table is None: |
| 1015 | return |
| 1016 | |
| 1017 | warmup_configs: set[SkinnyGemmConfig] = set() |
| 1018 | residual_warmup_configs: set[SkinnyGemmConfig] = set() |
| 1019 | for child in module.modules(): |
| 1020 | is_linear = ( |
| 1021 | isinstance(child, LinearBase) |
| 1022 | and type(child.quant_method) is UnquantizedLinearMethod |
| 1023 | ) |
| 1024 | # ParallelLMHead is a VocabParallelEmbedding subclass; embed_tokens is |
| 1025 | # the parent type, so isinstance already excludes it. |
| 1026 | is_head = ( |
| 1027 | isinstance(child, ParallelLMHead) |
| 1028 | and type(child.quant_method) is UnquantizedEmbeddingMethod |
| 1029 | ) |
| 1030 | if not (is_linear or is_head): |
| 1031 | continue |
| 1032 | weight = getattr(child, "weight", None) |
| 1033 | if weight is None or weight.dim() != 2: |
| 1034 | continue |
| 1035 | spec = table.get((weight.shape[0], weight.shape[1])) |
| 1036 | if spec is None: |
| 1037 | continue |
| 1038 | if is_linear: |
| 1039 | child.quant_method = KimiK3LowLatencyLinearMethod( |
| 1040 | _build_plan(spec), _build_residual_plan(spec) |
| 1041 | ) |
| 1042 | else: |
| 1043 | child.quant_method = KimiK3LowLatencyEmbeddingMethod(_build_plan(spec)) |
| 1044 | # Warm up only the configs measured for this module's local (N, K) so a |
| 1045 | # TP8 deployment does not compile TP4 configs and vice versa. |
| 1046 | warmup_configs.update(config for _, config in spec.cute_configs) |
| 1047 | residual_warmup_configs.update(config for _, config in spec.residual_configs) |
| 1048 | if _enable_kda_projection_overlap(module): |
| 1049 | from vllm.models.kimi_k3.nvidia.ops.cute_dsl.kda_skinny_gemm import ( |
| 1050 | kda_skinny_gemm, |
| 1051 | ) |
| 1052 | |
| 1053 | warmup_configs.update((*KDA_QKVG_CONFIGS.values(), KDA_M1_FAB_CONFIG)) |
| 1054 | kda_skinny_gemm.request_warmup( |
| 1055 | set(range(2, KDA_SKINNY_N_MAX_TOKENS + 1)), |
| 1056 | set(range(2, KDA_SKINNY_K_MAX_TOKENS + 1)), |
| 1057 | ) |
| 1058 | |
| 1059 | if shape_dynamic_skinny_gemm.is_available(): |
| 1060 | if warmup_configs: |
| 1061 | shape_dynamic_skinny_gemm.request_warmup_configs(dtype, warmup_configs) |
| 1062 | if residual_warmup_configs: |
| 1063 | shape_dynamic_skinny_gemm.request_warmup_configs( |
| 1064 | dtype, residual_warmup_configs, has_residual=True |
| 1065 | ) |