| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Based on: |
| 4 | Chen, L., Ye, Z., Wu, Y., Zhuo, D., Ceze, L., & Krishnamurthy, A. (2023). |
| 5 | Punica: Multi-Tenant LoRA Serving. |
| 6 | https://arxiv.org/abs/2310.18547 |
| 7 | """ |
| 8 | |
| 9 | from abc import ABC, abstractmethod |
| 10 | from typing import TYPE_CHECKING |
| 11 | |
| 12 | import torch |
| 13 | |
| 14 | from .utils import compute_meta, convert_mapping |
| 15 | |
| 16 | if TYPE_CHECKING: |
| 17 | # avoid circuit import |
| 18 | from vllm.lora.layers import LoRAMapping |
| 19 | |
| 20 | |
| 21 | class PunicaWrapperABC(ABC): |
| 22 | """PunicaWrapper ABC.""" |
| 23 | |
| 24 | @abstractmethod |
| 25 | def update_metadata( |
| 26 | self, |
| 27 | mapping: "LoRAMapping", |
| 28 | lora_index_to_id: list[int | None], |
| 29 | max_loras: int, |
| 30 | vocab_size: int, |
| 31 | **kwargs, |
| 32 | ) -> None: |
| 33 | """Update the lora-related metadata""" |
| 34 | raise NotImplementedError |
| 35 | |
| 36 | @abstractmethod |
| 37 | def add_shrink( |
| 38 | self, |
| 39 | y: tuple[torch.Tensor, ...] | torch.Tensor, |
| 40 | x: torch.Tensor, |
| 41 | lora_a_stacked: tuple[torch.Tensor, ...], |
| 42 | scale: float, |
| 43 | **kwargs, |
| 44 | ) -> torch.Tensor | None: |
| 45 | """Performs GEMM for multiple slices of lora_a.""" |
| 46 | raise NotImplementedError |
| 47 | |
| 48 | @abstractmethod |
| 49 | def add_expand( |
| 50 | self, |
| 51 | y: torch.Tensor, |
| 52 | x: tuple[torch.Tensor, ...] | torch.Tensor, |
| 53 | lora_b_stacked: tuple[torch.Tensor, ...], |
| 54 | output_slices: tuple[int, ...], |
| 55 | offset_start: int = 0, |
| 56 | add_inputs=True, |
| 57 | **kwargs, |
| 58 | ) -> torch.Tensor | None: |
| 59 | """Performs GEMM for multiple slices of lora_b.""" |
| 60 | raise NotImplementedError |
| 61 | |
| 62 | @abstractmethod |
| 63 | def add_lora_embedding( |
| 64 | self, |
| 65 | y: torch.Tensor, |
| 66 | x: torch.Tensor, |
| 67 | lora_b_stacked: torch.Tensor, |
| 68 | add_inputs: bool = True, |
| 69 | **kwargs, |
| 70 | ) -> torch.Tensor | None: |
| 71 | """Applies lora specifically for VocabParallelEmbeddingWithLoRA, |
| 72 | and this layer only requires the expand operation. |
| 73 | """ |
| 74 | raise NotImplementedError |
| 75 | |
| 76 | @abstractmethod |
| 77 | def add_lora_linear( |
| 78 | self, |
| 79 | y: torch.Tensor, |
| 80 | x: torch.Tensor, |
| 81 | lora_a_stacked: tuple[torch.Tensor, ...], |
| 82 | lora_b_stacked: tuple[torch.Tensor, ...], |
| 83 | scale: float, |
| 84 | output_slices: tuple[int, ...], |
| 85 | *, |
| 86 | buffer: tuple[torch.Tensor, ...] | None = None, |
| 87 | **kwargs, |
| 88 | ) -> torch.Tensor | None: |
| 89 | """Applicable to linear-related lora.""" |
| 90 | raise NotImplementedError |
| 91 | |
| 92 | @abstractmethod |
| 93 | def add_lora_logits( |
| 94 | self, |
| 95 | y: torch.Tensor, |
| 96 | x: torch.Tensor, |
| 97 | lora_a_stacked: torch.Tensor, |
| 98 | lora_b_stacked: torch.Tensor, |
| 99 | scale, |
| 100 | *, |
| 101 | buffer: torch.Tensor | None = None, |
| 102 | **kwargs, |
| 103 | ) -> torch.Tensor | None: |
| 104 | """Applies lora specifically for LogitsProcessorWithLoRA.""" |
| 105 | raise NotImplementedError |
| 106 | |
| 107 | @abstractmethod |
| 108 | def apply_lora_full_linear( |
| 109 | self, |
| 110 | y: torch.Tensor, |
| 111 | x: torch.Tensor, |
| 112 | weight_stacked: torch.Tensor, |
| 113 | bias_stacked: torch.Tensor, |
| 114 | module_enabled: torch.Tensor, |
| 115 | ) -> torch.Tensor | None: |
| 116 | """Apply request-routed full linear weights to selected rows.""" |
| 117 | raise NotImplementedError |
| 118 | |
| 119 | |
| 120 | class PunicaWrapperBase(PunicaWrapperABC): |
| 121 | """PunicaWrapperBase is designed to manage and provide metadata for the punica |
| 122 | kernel. The main function is to maintain the state information for |
| 123 | Multi-LoRA, and to provide the interface for the punica. |
| 124 | """ |
| 125 | |
| 126 | def __init__( |
| 127 | self, |
| 128 | max_num_batched_tokens: int, |
| 129 | max_batches: int, |
| 130 | device: torch.device | str, |
| 131 | **kwargs, |
| 132 | ): |
| 133 | self._token_lora_indices = torch.empty( |
| 134 | max_num_batched_tokens, dtype=torch.long, device=device |
| 135 | ) |
| 136 | self._sampler_indices = torch.empty( |
| 137 | max_num_batched_tokens, dtype=torch.long, device=device |
| 138 | ) |
| 139 | self._sampler_indices_padded = torch.empty( |
| 140 | max_num_batched_tokens, dtype=torch.long, device=device |
| 141 | ) |
| 142 | self._embeddings_indices = torch.empty( |
| 143 | max_num_batched_tokens, dtype=torch.long, device=device |
| 144 | ) |
| 145 | |
| 146 | # 4 is the number of indices tensors. |
| 147 | # base_indices, sampler_indices, sampler_indices_padded, |
| 148 | # embeddings_indices |
| 149 | self.indices_len: list[int | None] = [None] * 4 |
| 150 | # these attributes are the information required for sgmv kernel |
| 151 | self._seq_start_locs = torch.empty(max_batches, dtype=torch.long, device=device) |
| 152 | self._seq_lengths = torch.empty(max_batches, dtype=torch.long, device=device) |
| 153 | self._lora_indices_per_batch = torch.empty( |
| 154 | max_batches, dtype=torch.long, device=device |
| 155 | ) |
| 156 | self.device: torch.device = device |
| 157 | self.max_length: int = 0 |
| 158 | self.token_nums: int = 0 |
| 159 | self.batch_size: int = -1 |
| 160 | self.is_prefill = False |
| 161 | self.no_lora = False |
| 162 | |
| 163 | def _update_base_metadata( |
| 164 | self, |
| 165 | mapping: "LoRAMapping", |
| 166 | lora_index_to_id: list[int | None], |
| 167 | max_loras: int, |
| 168 | vocab_size: int, |
| 169 | ): |
| 170 | ( |
| 171 | base_indices, |
| 172 | sampler_indices, |
| 173 | sampler_indices_padded, |
| 174 | embeddings_indices, |
| 175 | indices_len, |
| 176 | ) = convert_mapping( |
| 177 | mapping, |
| 178 | lora_index_to_id, |
| 179 | max_loras, |
| 180 | vocab_size, |
| 181 | self.device, |
| 182 | ) |
| 183 | self._token_lora_indices[: base_indices.shape[0]].copy_(base_indices) |
| 184 | self._sampler_indices[: sampler_indices.shape[0]].copy_(sampler_indices) |
| 185 | self._sampler_indices_padded[: sampler_indices_padded.shape[0]].copy_( |
| 186 | sampler_indices_padded |
| 187 | ) |
| 188 | self._embeddings_indices[: embeddings_indices.shape[0]].copy_( |
| 189 | embeddings_indices |
| 190 | ) |
| 191 | |
| 192 | self.indices_len[:] = indices_len |
| 193 | |
| 194 | def _update_prefill_metadata(self, token_lora_tensor: torch.Tensor) -> None: |
| 195 | ( |
| 196 | b_seq_start_tensor, |
| 197 | seq_length_tensor, |
| 198 | lora_indices_tensor, |
| 199 | batch_size, |
| 200 | max_length, |
| 201 | token_nums, |
| 202 | no_lora, |
| 203 | ) = compute_meta(token_lora_tensor) |
| 204 | |
| 205 | self._seq_start_locs[: b_seq_start_tensor.shape[0]].copy_(b_seq_start_tensor) |
| 206 | self._seq_lengths[: seq_length_tensor.shape[0]].copy_(seq_length_tensor) |
| 207 | self._lora_indices_per_batch[: lora_indices_tensor.shape[0]].copy_( |
| 208 | lora_indices_tensor |
| 209 | ) |
| 210 | self.batch_size = batch_size |
| 211 | self.max_length = max_length |
| 212 | self.token_nums = token_nums |
| 213 | self.no_lora = no_lora |
| 214 | |
| 215 | def _select_full_linear_output( |
| 216 | self, |
| 217 | y: torch.Tensor, |
| 218 | adapter_y: torch.Tensor, |
| 219 | bias_stacked: torch.Tensor, |
| 220 | module_enabled: torch.Tensor, |
| 221 | ) -> torch.Tensor: |
| 222 | indices = self.sampler_indices |
| 223 | assert indices.size(0) == y.size(0), ( |
| 224 | "Full linear rows do not match LoRA request mapping" |
| 225 | ) |
| 226 | safe_indices = indices.clamp_min(0) |
| 227 | use_full = (indices >= 0) & module_enabled[safe_indices] |
| 228 | adapter_y = adapter_y.to(y.dtype) + bias_stacked[safe_indices].to(y.dtype) |
| 229 | return torch.where(use_full.unsqueeze(-1), adapter_y, y) |
| 230 | |
| 231 | @property |
| 232 | def prefill_metadata( |
| 233 | self, |
| 234 | ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, int, int, int]: |
| 235 | """This property provides a convenient way to access the necessary |
| 236 | metadata for prefill-related kernel computations. |
| 237 | 1. seq_start_locs: Tensor of sequence start positions. |
| 238 | 2. seq_lengths: Tensor of sequence lengths. |
| 239 | 3. lora_indices_per_batch: Tensor of lora indices, and an index of |
| 240 | -1 means no lora should be applied. |
| 241 | 4. batch_size: Batch size after clustering identical lora indices. |
| 242 | 5. max_length: The maximum sequence length in the batch. |
| 243 | 6. token_nums: The token numbers in the batch. |
| 244 | """ |
| 245 | return ( |
| 246 | self._seq_start_locs[: self.batch_size], |
| 247 | self._seq_lengths[: self.batch_size], |
| 248 | self._lora_indices_per_batch[: self.batch_size], |
| 249 | self.batch_size, |
| 250 | self.max_length, |
| 251 | self.token_nums, |
| 252 | ) |
| 253 | |
| 254 | @property |
| 255 | def token_lora_indices(self) -> torch.Tensor: |
| 256 | """This property provides the lora indices corresponding to each token |
| 257 | in the batch. An index of -1 means no lora should be applied. |
| 258 | """ |
| 259 | token_lora_len = self.indices_len[0] |
| 260 | return self._token_lora_indices[:token_lora_len] |
| 261 | |
| 262 | @property |
| 263 | def sampler_indices(self) -> torch.Tensor: |
| 264 | """This property is used to access the lora indices specifically for |
| 265 | LogitsProcessorWithLoRA. |
| 266 | """ |
| 267 | sampler_indices_len = self.indices_len[1] |
| 268 | return self._sampler_indices[:sampler_indices_len] |
| 269 | |
| 270 | @property |
| 271 | def sampler_indices_padded(self) -> torch.Tensor: |
| 272 | """This property provides access to padded sampler indices.""" |
| 273 | indices_padded_len = self.indices_len[2] |
| 274 | return self._sampler_indices_padded[:indices_padded_len] |
| 275 | |
| 276 | def update_metadata( |
| 277 | self, |
| 278 | mapping: "LoRAMapping", |
| 279 | lora_index_to_id: list[int | None], |
| 280 | max_loras: int, |
| 281 | vocab_size: int, |
| 282 | **kwargs, |
| 283 | ): |
| 284 | self._update_base_metadata(mapping, lora_index_to_id, max_loras, vocab_size) |
| 285 | |
| 286 | if mapping.is_prefill: |
| 287 | # Update metadata required for prefill-related operators. |
| 288 | self._update_prefill_metadata(self.token_lora_indices) |
| 289 | self.is_prefill = True |
| 290 | else: |
| 291 | self.is_prefill = False |
| 292 | |
| 293 | @abstractmethod |
| 294 | def add_shrink( |
| 295 | self, |
| 296 | y: tuple[torch.Tensor, ...] | torch.Tensor, |
| 297 | x: torch.Tensor, |
| 298 | lora_a_stacked: tuple[torch.Tensor, ...], |
| 299 | scale: float, |
| 300 | **kwargs, |
| 301 | ) -> torch.Tensor | None: |
| 302 | """Performs GEMM for multiple slices of lora_a. |
| 303 | |
| 304 | Semantics: |
| 305 | for i in range(len(lora_a_stacked)): |
| 306 | y[i] += (x @ lora_a_stacked[i]) * scale |
| 307 | |
| 308 | Args: |
| 309 | y (Union[tuple[torch.Tensor, ...], torch.Tensor]): Output tensors |
| 310 | x (torch.Tensor): Input tensor |
| 311 | lora_a_stacked (tuple[torch.Tensor, ...]): lora_a's weights |
| 312 | scale (float): Scaling factor for the operation |
| 313 | |
| 314 | """ |
| 315 | # TODO: implement it based on torch ops |
| 316 | raise NotImplementedError |
| 317 | |
| 318 | @abstractmethod |
| 319 | def add_expand( |
| 320 | self, |
| 321 | y: torch.Tensor, |
| 322 | x: tuple[torch.Tensor, ...] | torch.Tensor, |
| 323 | lora_b_stacked: tuple[torch.Tensor, ...], |
| 324 | output_slices: tuple[int, ...], |
| 325 | offset_start: int = 0, |
| 326 | add_inputs=True, |
| 327 | **kwargs, |
| 328 | ) -> torch.Tensor | None: |
| 329 | """Performs GEMM for multiple slices of lora_b. |
| 330 | |
| 331 | Semantics: |
| 332 | offset = offset_start |
| 333 | for i in range(len(lora_b_stacked)): |
| 334 | slice = output_slices[i] |
| 335 | y[:, offset:offset+slice] += x[i] @ lora_b_stacked[i] |
| 336 | offset += slice |
| 337 | |
| 338 | Args: |
| 339 | y (torch.Tensor): Output tensor. |
| 340 | x (Union[tuple[torch.Tensor, ...], torch.Tensor]): Input tensors |
| 341 | lora_b_stacked (tuple[torch.Tensor, ...]): lora_b's weight |
| 342 | output_slices (tuple[int, ...]): Every slice's size |
| 343 | offset_start (int): The starting position of y, defaults to 0 |
| 344 | add_inputs (bool): Defaults to True. |
| 345 | |
| 346 | """ |
| 347 | # TODO: implement it based on torch ops |
| 348 | raise NotImplementedError |
| 349 | |
| 350 | @abstractmethod |
| 351 | def add_lora_embedding( |
| 352 | self, |
| 353 | y: torch.Tensor, |
| 354 | x: torch.Tensor, |
| 355 | lora_b_stacked: torch.Tensor, |
| 356 | add_inputs: bool = True, |
| 357 | **kwargs, |
| 358 | ) -> torch.Tensor | None: |
| 359 | """Applies lora specifically for VocabParallelEmbeddingWithLoRA. |
| 360 | and this layer only requires the expand operation. |
| 361 | Semantics: |
| 362 | y += x @ lora_b_stacked |
| 363 | |
| 364 | Args: |
| 365 | y (torch.Tensor): Output tensor. |
| 366 | x (torch.Tensor): Input tensor. |
| 367 | lora_b_stacked (torch.Tensor): lora_b's weights. |
| 368 | add_inputs (bool): Default to True. |
| 369 | |
| 370 | """ |
| 371 | # TODO: implement it based on torch ops |
| 372 | raise NotImplementedError |
| 373 | |
| 374 | @abstractmethod |
| 375 | def add_lora_linear( |
| 376 | self, |
| 377 | y: torch.Tensor, |
| 378 | x: torch.Tensor, |
| 379 | lora_a_stacked: tuple[torch.Tensor, ...], |
| 380 | lora_b_stacked: tuple[torch.Tensor, ...], |
| 381 | scale: float, |
| 382 | output_slices: tuple[int, ...], |
| 383 | *, |
| 384 | buffer: tuple[torch.Tensor, ...] | None = None, |
| 385 | **kwargs, |
| 386 | ) -> torch.Tensor | None: |
| 387 | """Applicable to linear-related lora. |
| 388 | |
| 389 | Semantics: |
| 390 | for i in range(len(lora_a_stacked)): |
| 391 | y[i] += ( |
| 392 | x[i].unsqueeze(0) |
| 393 | @ lora_a_stacked[indices[i], layer_idx, :, :] |
| 394 | @ lora_b_stacked[indices[i], layer_idx, :, :] |
| 395 | * scale |
| 396 | ).squeeze(0) |
| 397 | |
| 398 | Args: |
| 399 | y (torch.Tensor): Output tensor. Will be changed in-place. |
| 400 | x (torch.Tensor): Input tensor |
| 401 | lora_a_stacked (tuple[torch.Tensor, ...]): lora_a's weight. |
| 402 | lora_b_stacked (tuple[torch.Tensor, ...]): lora_b's weight. |
| 403 | scale (float): Scaling factor. |
| 404 | output_slices (tuple[int, ...]): Every slice's size. |
| 405 | buffer (Optional[tuple[torch.Tensor, ...]]): Defaults to None. |
| 406 | |
| 407 | """ |
| 408 | # TODO: implement it based on torch ops |
| 409 | raise NotImplementedError |
| 410 | |
| 411 | @abstractmethod |
| 412 | def add_lora_logits( |
| 413 | self, |
| 414 | y: torch.Tensor, |
| 415 | x: torch.Tensor, |
| 416 | lora_a_stacked: torch.Tensor, |
| 417 | lora_b_stacked: torch.Tensor, |
| 418 | scale, |
| 419 | *, |
| 420 | buffer: torch.Tensor | None = None, |
| 421 | **kwargs, |
| 422 | ) -> torch.Tensor | None: |
| 423 | """Applies lora specifically for LogitsProcessorWithLoRA. |
| 424 | |
| 425 | Semantics: |
| 426 | buffer = (x @ lora_a_stacked) * scale |
| 427 | y += buffer @ lora_b_stacked |
| 428 | |
| 429 | Args: |
| 430 | y (torch.Tensor): Output tensor. |
| 431 | x (torch.Tensor): Input tensor. |
| 432 | lora_a_stacked (torch.Tensor): lora_a's weights. |
| 433 | lora_b_stacked (torch.Tensor):lora_b's weights. |
| 434 | scale (float): Scaling factor. |
| 435 | buffer (Optional[torch.Tensor]):Default to None. |
| 436 | |
| 437 | """ |
| 438 | # TODO: implement it based on torch ops |
| 439 | raise NotImplementedError |
| 440 | |
| 441 | def moe_lora_align_block_size( |
| 442 | self, |
| 443 | topk_ids: torch.Tensor, |
| 444 | num_tokens: int, |
| 445 | block_size: int, |
| 446 | num_experts: int, |
| 447 | max_loras: int, |
| 448 | adapter_enabled: torch.Tensor, |
| 449 | expert_map: torch.Tensor | None = None, |
| 450 | pad_sorted_ids: bool = False, |
| 451 | naive_block_assignment: bool = False, |
| 452 | ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: |
| 453 | """Aligns tokens and experts into block-sized chunks for LoRA-based |
| 454 | mixture-of-experts (MoE) execution. |
| 455 | """ |
| 456 | # TODO: implement it based on torch ops |
| 457 | raise NotImplementedError |
| 458 | |
| 459 | def add_lora_fused_moe( |
| 460 | self, |
| 461 | y: torch.Tensor, |
| 462 | x: torch.Tensor, |
| 463 | lora_a_stacked: tuple[torch.Tensor, ...], |
| 464 | lora_b_stacked: tuple[torch.Tensor, ...], |
| 465 | topk_weights: torch.Tensor, |
| 466 | sorted_token_ids: torch.Tensor | None, |
| 467 | expert_ids: torch.Tensor, |
| 468 | num_tokens_post_padded: torch.Tensor | None, |
| 469 | max_lora_rank: int, |
| 470 | top_k_num: int, |
| 471 | shrink_config, |
| 472 | expand_config, |
| 473 | adapter_enabled: torch.Tensor, |
| 474 | mul_routed_weight=False, |
| 475 | fully_sharded: bool = False, |
| 476 | offset: int = 0, |
| 477 | token_lora_mapping: torch.Tensor | None = None, |
| 478 | ): |
| 479 | """Performs a fused forward computation for LoRA of |
| 480 | Mixture-of-Experts (MoE) layer. |
| 481 | """ |
| 482 | # TODO: implement it based on torch ops |
| 483 | raise NotImplementedError |
| 484 | |
| 485 | def add_lora_w13( |
| 486 | self, |
| 487 | y: torch.Tensor, |
| 488 | x: torch.Tensor, |
| 489 | lora_a_stacked: tuple[torch.Tensor, ...], |
| 490 | lora_b_stacked: tuple[torch.Tensor, ...], |
| 491 | topk_ids: torch.Tensor, |
| 492 | topk_weights: torch.Tensor, |
| 493 | expert_map: torch.Tensor | None, |
| 494 | w1: torch.Tensor, |
| 495 | w2: torch.Tensor, |
| 496 | num_tokens: int, |
| 497 | top_k_num: int, |
| 498 | max_loras: int, |
| 499 | adapter_enabled: torch.Tensor, |
| 500 | local_num_experts: int, |
| 501 | top_k: int, |
| 502 | num_slices: int, |
| 503 | fully_sharded: bool, |
| 504 | use_tuned_config: bool, |
| 505 | add_inputs: bool = True, |
| 506 | token_lora_mapping: torch.Tensor | None = None, |
| 507 | ) -> tuple[ |
| 508 | torch.Tensor | None, |
| 509 | torch.Tensor | None, |
| 510 | torch.Tensor | None, |
| 511 | torch.Tensor | None, |
| 512 | ]: |
| 513 | """Apply w13 LoRA to y (intermediate_cache1) in-place before activation. |
| 514 | |
| 515 | When `token_lora_mapping` is provided it overrides the punica_wrapper's |
| 516 | global mapping — used by EP+LoRA to pass the per-rank-local mapping |
| 517 | after all-to-all dispatch. |
| 518 | |
| 519 | Returns (sorted_token_ids_lora, expert_ids_lora, |
| 520 | num_tokens_post_padded_lora, token_lora_mapping) |
| 521 | for reuse by add_lora_w2. |
| 522 | """ |
| 523 | raise NotImplementedError |
| 524 | |
| 525 | def add_lora_w2( |
| 526 | self, |
| 527 | y: torch.Tensor, |
| 528 | x: torch.Tensor, |
| 529 | lora_a_stacked: tuple[torch.Tensor, ...], |
| 530 | lora_b_stacked: tuple[torch.Tensor, ...], |
| 531 | topk_weights: torch.Tensor, |
| 532 | sorted_token_ids_lora: torch.Tensor | None, |
| 533 | expert_ids_lora: torch.Tensor | None, |
| 534 | num_tokens_post_padded_lora: torch.Tensor | None, |
| 535 | token_lora_mapping: torch.Tensor | None, |
| 536 | num_tokens: int, |
| 537 | w1: torch.Tensor, |
| 538 | w2: torch.Tensor, |
| 539 | top_k_num: int, |
| 540 | max_loras: int, |
| 541 | adapter_enabled: torch.Tensor, |
| 542 | top_k: int, |
| 543 | fully_sharded: bool, |
| 544 | tp_rank: int, |
| 545 | use_tuned_config: bool, |
| 546 | add_inputs: bool = True, |
| 547 | ) -> None: |
| 548 | """Apply w2 LoRA to y (intermediate_cache3) in-place before moe_sum. |
| 549 | |
| 550 | Reuses routing tensors returned by add_lora_w13. |
| 551 | """ |
| 552 | raise NotImplementedError |