layers.py 47.8 KB
Newer Older
1
2
3
# pylint: disable=unused-argument
import math
from dataclasses import dataclass
4
from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
5
6
7
8
9
10

import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import PretrainedConfig

11
from vllm.adapter_commons.layers import AdapterMapping
12
from vllm.config import LoRAConfig
13
14
15
16
17
18
from vllm.distributed import (get_tensor_model_parallel_rank,
                              get_tensor_model_parallel_world_size,
                              split_tensor_along_last_dim,
                              tensor_model_parallel_all_gather,
                              tensor_model_parallel_all_reduce,
                              tensor_model_parallel_gather)
19
from vllm.distributed.utils import divide
20
21
from vllm.lora.punica import add_lora, add_lora_slice, bgmv
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
22
                                               MergedColumnParallelLinear,
23
                                               QKVParallelLinear,
24
                                               RowParallelLinear)
25
from vllm.model_executor.layers.logits_processor import LogitsProcessor
26
27
from vllm.model_executor.layers.rotary_embedding import (
    LinearScalingRotaryEmbedding, RotaryEmbedding)
28
from vllm.model_executor.layers.vocab_parallel_embedding import (
29
    VocabParallelEmbedding)
30
31
32
33
34

if TYPE_CHECKING:
    pass


35
36
37
def _get_lora_device(base_layer: nn.Module) -> torch.device:
    # code borrowed from https://github.com/fmmoret/vllm/blob/fm-support-lora-on-quantized-models/vllm/lora/layers.py#L34
    """Returns the device for where to place the LoRA tensors."""
Jee Li's avatar
Jee Li committed
38
    # unquantizedLinear
39
40
    if hasattr(base_layer, "weight"):
        return base_layer.weight.device
Jee Li's avatar
Jee Li committed
41
42
43
44
45
46
47
48
    # GPTQ/AWQ/SqueezeLLM
    elif hasattr(base_layer, "qweight"):
        return base_layer.qweight.device
    # marlin
    elif hasattr(base_layer, "B"):
        return base_layer.B.device
    else:
        raise ValueError(f"Unsupported base layer: {base_layer}")
49
50


51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
def _not_fully_sharded_can_replace(can_replace):
    """
    decorator which adds the condition of not using fully sharded loras
    intended to wrap can_replace_layer()
    """

    def dec(*args, **kwargs):
        decorate = kwargs.pop('decorate') if 'decorate' in kwargs else True
        condition = (not kwargs['lora_config'].fully_sharded_loras
                     if decorate else True)
        return can_replace(*args, **kwargs) and condition

    return dec


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
def _apply_lora(
    x: torch.Tensor,
    lora_a_stacked: torch.Tensor,
    lora_b_stacked: torch.Tensor,
    indices: torch.Tensor,
    output: torch.Tensor,
):
    """Applies lora to each input.

    This method applies all loras to each input. It uses the
    indices vector to determine which lora yields the
    correct output. An index of -1 means no lora should be
    applied. This method adds the final lora results to the
    output.

    Input shapes:
        x:               (batch_size, hidden_dim)
        lora_a_stacked:  (num_loras, lora_rank, hidden_dim)
        lora_b_stacked:  (num_loras, output_dim, lora_rank)
        indices:         (batch_size)
        output:          (batch_size, output_dim)
    """
    org_output = output
    x = x.view(-1, x.shape[-1])
    output = output.view(-1, output.shape[-1])
    indices = indices.view(-1)
    add_lora(output, x, lora_a_stacked, lora_b_stacked, indices, 0, 1.0)
    return output.view_as(org_output)


def _apply_lora_packed_nslice(
    x: torch.Tensor,
    lora_a_stacked: Tuple[torch.Tensor, torch.Tensor, torch.Tensor],
    lora_b_stacked: Tuple[torch.Tensor, torch.Tensor, torch.Tensor],
    indices: torch.Tensor,
    output: torch.Tensor,
    output_slices: Tuple[int, ...],
):
    """Applies lora to each input.

    This method applies all loras to each input. It uses the
    indices vector to determine which lora yields the
    correct output. An index of -1 means no lora should be
    applied. This method adds the final lora results to the
    output.

    This method is used for layers that are composed of multiple sublayers
    (slices) packed together.

    Input shapes:
        x:                 (batch_size, hidden_dim)
        lora_a_stacked:    3 element tuple of (num_loras, lora_rank, hidden_dim)
        lora_b_stacked:    3 element tuple of (num_loras, output_dim, lora_rank)
        indices:           (batch_size)
        output:            (batch_size, q_slice_size + 2*kv_slice_size)
121
122
        output_slices:     n-1 element tuple of (slice_size...),
                           where n is number of slices
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
    """
    org_output = output
    x = x.view(-1, x.shape[-1])
    output = output.view(-1, output.shape[-1])
    indices = indices.view(-1)
    offset_left = 0
    for slice_idx in range(len(output_slices)):
        add_lora_slice(output, x, lora_a_stacked[slice_idx],
                       lora_b_stacked[slice_idx], indices, 0, 1.0, offset_left,
                       output_slices[slice_idx])
        offset_left += output_slices[slice_idx]
    return output.view_as(org_output)


@dataclass
138
139
class LoRAMapping(AdapterMapping):
    pass
140
141
142
143


class BaseLayerWithLoRA(nn.Module):

144
145
146
    def slice_lora_a(
        self, lora_a: Union[torch.Tensor, List[Union[torch.Tensor, None]]]
    ) -> Union[torch.Tensor, List[Union[torch.Tensor, None]]]:
147
148
149
        """Slice lora a if splitting for tensor parallelism."""
        ...

150
151
152
    def slice_lora_b(
        self, lora_b: Union[torch.Tensor, List[Union[torch.Tensor, None]]]
    ) -> Union[torch.Tensor, List[Union[torch.Tensor, None]]]:
153
154
155
        """Slice lora b if splitting with tensor parallelism."""
        ...

156
157
158
159
160
    def create_lora_weights(
            self,
            max_loras: int,
            lora_config: LoRAConfig,
            model_config: Optional[PretrainedConfig] = None) -> None:
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
        """Initializes lora matrices."""
        ...

    def reset_lora(self, index: int):
        """Resets the lora weights at index back to 0."""
        ...

    def set_lora(
        self,
        index: int,
        lora_a: torch.Tensor,
        lora_b: torch.Tensor,
        embeddings_tensor: Optional[torch.Tensor],
    ):
        """Overwrites lora tensors at index."""
        ...

    def set_mapping(
        self,
        base_indices: torch.Tensor,
        sampler_indices: torch.Tensor,
        sampler_indices_padded: torch.Tensor,
        embeddings_indices: torch.Tensor,
184
        long_lora_indices: torch.Tensor,
185
186
187
188
189
        indices_len: List[int],
    ):
        """Sets the mapping indices."""
        ...

190
191
192
193
194
195
196
    @classmethod
    def can_replace_layer(cls, source_layer: nn.Module,
                          lora_config: LoRAConfig, packed_modules_list: List,
                          model_config: Optional[PretrainedConfig]) -> bool:
        """Returns True if the layer can be replaced by this LoRA layer."""
        raise NotImplementedError

197
198
199
200
201
202

class VocabParallelEmbeddingWithLoRA(BaseLayerWithLoRA):

    def __init__(self, base_layer: VocabParallelEmbedding) -> None:
        super().__init__()
        self.base_layer = base_layer
203
204
        self.embeddings_slice: Optional[Tuple[int, int]]
        self.embeddings_weights: Optional[torch.Tensor]
205
206
207
208
209
210
211

    def create_lora_weights(
            self,
            max_loras: int,
            lora_config: LoRAConfig,
            model_config: Optional[PretrainedConfig] = None) -> None:

212
        if self.base_layer.num_added_embeddings_per_partition > 0:
213
            # We can start adding lora weights
214
215
216
217
218
219
220
221
222
223
224
            self.embeddings_weights = self.base_layer.weight.data[
                self.base_layer.num_org_embeddings_per_partition:self.
                base_layer.num_org_embeddings_per_partition +
                self.base_layer.num_added_embeddings_per_partition]
            self.embeddings_slice = (
                self.base_layer.shard_indices.added_vocab_start_index -
                self.base_layer.org_vocab_size,
                self.base_layer.shard_indices.added_vocab_end_index -
                self.base_layer.org_vocab_size)
            self.base_layer.weight.data[
                self.base_layer.num_org_embeddings_per_partition:].fill_(0)
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
        else:
            self.embeddings_slice = None
            self.embeddings_weights = None

        self.embeddings_tensors = torch.zeros(
            (
                max_loras,
                lora_config.lora_extra_vocab_size,
                self.base_layer.embedding_dim,
            ),
            dtype=self.base_layer.weight.dtype,
            device=self.base_layer.weight.device,
        )
        self.lora_a_stacked = torch.zeros(
            (
                max_loras,
                self.base_layer.org_vocab_size +
                lora_config.lora_extra_vocab_size,
                lora_config.max_lora_rank,
            ),
            dtype=lora_config.lora_dtype,
            device=self.base_layer.weight.device,
        )
        self.lora_b_stacked = torch.zeros(
            (
                max_loras,
                1,
                self.base_layer.embedding_dim,
                lora_config.max_lora_rank,
            ),
            dtype=lora_config.lora_dtype,
            device=self.base_layer.weight.device,
        )
        self.lora_a_stacked_2d = self.lora_a_stacked.view(
            self.lora_a_stacked.shape[0] * self.lora_a_stacked.shape[1],
            self.lora_a_stacked.shape[2],
        )
262
263
264
265
        # Lazily initialized.
        self.indices: torch.Tensor
        self.indices_len: List[int]
        self.embeddings_indices: torch.Tensor
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

    def reset_lora(self, index: int):
        self.lora_a_stacked[index] = 0
        self.lora_b_stacked[index] = 0
        self.embeddings_tensors[index] = 0

    def set_lora(
        self,
        index: int,
        lora_a: torch.Tensor,
        lora_b: torch.Tensor,
        embeddings_tensor: Optional[torch.Tensor],
    ):
        self.reset_lora(index)
        self.lora_a_stacked[index, :lora_a.shape[0], :lora_a.shape[1]].copy_(
            lora_a, non_blocking=True)
        self.lora_b_stacked[index,
                            0, :lora_b.shape[1], :lora_b.shape[0]].copy_(
                                lora_b.T, non_blocking=True)
        if embeddings_tensor is not None:
            self.embeddings_tensors[
                index, :embeddings_tensor.shape[0], :embeddings_tensor.
                shape[1]].copy_(embeddings_tensor, non_blocking=True)
            if self.embeddings_slice is not None:
                # TODO(yard1): Optimize this copy, we don't need to copy
                # everything, just the modified part
                embeddings = self.embeddings_tensors.view(
                    self.embeddings_tensors.shape[0] *
                    self.embeddings_tensors.shape[1],
                    self.embeddings_tensors.shape[2]
                )[self.embeddings_slice[0]:self.embeddings_slice[1]]
297
                assert self.embeddings_weights is not None
298
299
300
301
302
303
304
305
                self.embeddings_weights[:embeddings.shape[0]].copy_(embeddings)

    def set_mapping(
        self,
        base_indices: torch.Tensor,
        sampler_indices: torch.Tensor,
        sampler_indices_padded: torch.Tensor,
        embeddings_indices: torch.Tensor,
306
        long_lora_indices: torch.Tensor,
307
308
309
310
311
312
313
314
        indices_len: List[int],
    ):
        self.indices = base_indices
        self.embeddings_indices = embeddings_indices
        self.indices_len = indices_len

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        added_tokens_mask = x > self.base_layer.org_vocab_size - 1
315
316
        embedding_len = self.indices_len[3]
        indices = self.embeddings_indices[1][:embedding_len].view_as(x)
317
318
319
320
        full_lora_a_embeddings = F.embedding(
            x + indices,
            self.lora_a_stacked_2d,
        )
321
        indices = self.embeddings_indices[0][:embedding_len].view_as(x)
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
        full_output = self.base_layer.forward(
            x.add_(indices * added_tokens_mask))

        full_output_org = full_output
        if full_output.ndim == 3:
            full_output = full_output.view(
                full_output.shape[0] * full_output.shape[1], -1)
        if full_lora_a_embeddings.ndim == 3:
            full_lora_a_embeddings = full_lora_a_embeddings.view(
                full_lora_a_embeddings.shape[0] *
                full_lora_a_embeddings.shape[1], -1)
        bgmv(full_output, full_lora_a_embeddings, self.lora_b_stacked,
             self.indices[:self.indices_len[0]], 0, 1.0)
        return full_output.view_as(full_output_org)

337
338
339
340
341
342
    @classmethod
    def can_replace_layer(cls, source_layer: nn.Module,
                          lora_config: LoRAConfig, packed_modules_list: List,
                          model_config: Optional[PretrainedConfig]) -> bool:
        return type(source_layer) is VocabParallelEmbedding

343
344

class ColumnParallelLinearWithLoRA(BaseLayerWithLoRA):
345
346
347
348
349
    """
    LoRA on top of ColumnParallelLinear layer.
    
    LoRA B is sliced for tensor parallelism.
    """
350
351
352
353

    def __init__(self, base_layer: ColumnParallelLinear) -> None:
        super().__init__()
        self.base_layer = base_layer
354
        self.tp_size = get_tensor_model_parallel_world_size()
355
356
357
        self.input_size = self.base_layer.input_size
        self.output_size = self.base_layer.output_size_per_partition
        self.device = _get_lora_device(self.base_layer)
358
359
360
361
362
363

    def create_lora_weights(
            self,
            max_loras: int,
            lora_config: LoRAConfig,
            model_config: Optional[PretrainedConfig] = None) -> None:
364
365
366
367
368
        self.lora_config = lora_config
        self.tp_size = get_tensor_model_parallel_world_size()
        lora_a_output_size_per_partition = (
            lora_config.max_lora_rank if not lora_config.fully_sharded_loras
            else divide(lora_config.max_lora_rank, self.tp_size))
369
370
371
        self.lora_a_stacked = torch.zeros(
            max_loras,
            1,
372
            lora_a_output_size_per_partition,
373
            self.input_size,
374
            dtype=lora_config.lora_dtype,
375
            device=self.device,
376
377
378
379
        )
        self.lora_b_stacked = torch.zeros(
            max_loras,
            1,
380
            self.output_size,
381
382
            lora_config.max_lora_rank,
            dtype=lora_config.lora_dtype,
383
            device=self.device,
384
        )
385
        self.output_dim = self.lora_b_stacked.shape[2]
386

387
388
389
390
        # lazily initialized.
        self.indices: torch.Tensor
        self.indices_len: List[int]

391
392
393
394
    def reset_lora(self, index: int):
        self.lora_a_stacked[index] = 0
        self.lora_b_stacked[index] = 0

395
396
397
398
399
400
401
402
403
404
405
    def slice_lora_a(self, lora_a: torch.Tensor) -> torch.Tensor:
        return lora_a

    def slice_lora_b(self, lora_b: torch.Tensor) -> torch.Tensor:
        tensor_model_parallel_rank = get_tensor_model_parallel_rank()
        shard_size = self.output_dim
        start_idx = tensor_model_parallel_rank * shard_size
        end_idx = (tensor_model_parallel_rank + 1) * shard_size
        lora_b = lora_b[:, start_idx:end_idx]
        return lora_b

406
407
408
409
410
411
412
413
    def set_lora(
        self,
        index: int,
        lora_a: torch.Tensor,
        lora_b: torch.Tensor,
        embeddings_tensor: Optional[torch.Tensor],
    ):
        self.reset_lora(index)
414

415
        if self.tp_size > 1:
416
417
418
            lora_a = self.slice_lora_a(lora_a)
            lora_b = self.slice_lora_b(lora_b)

419
420
421
422
423
424
425
426
427
428
429
430
431
        self.lora_a_stacked[index,
                            0, :lora_a.shape[1], :lora_a.shape[0]].copy_(
                                lora_a.T, non_blocking=True)
        self.lora_b_stacked[index,
                            0, :lora_b.shape[1], :lora_b.shape[0]].copy_(
                                lora_b.T, non_blocking=True)

    def set_mapping(
        self,
        base_indices: torch.Tensor,
        sampler_indices: torch.Tensor,
        sampler_indices_padded: torch.Tensor,
        embeddings_indices: torch.Tensor,
432
        long_lora_indices: torch.Tensor,
433
434
435
436
437
        indices_len: List[int],
    ):
        self.indices = base_indices
        self.indices_len = indices_len

438
439
440
    def apply(self, x: torch.Tensor,
              bias: Optional[torch.Tensor]) -> torch.Tensor:
        output = self.base_layer.quant_method.apply(self.base_layer, x, bias)
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
        _apply_lora(
            x,
            self.lora_a_stacked,
            self.lora_b_stacked,
            self.indices[:self.indices_len[0]],
            output,
        )
        return output

    def forward(self, input_):
        """Forward of ColumnParallelLinear

        Args:
            input_: Tensor whose last dimension is `input_size`.

        Returns:
            - output
            - bias
        """
        bias = (self.base_layer.bias
                if not self.base_layer.skip_bias_add else None)

        # Matrix multiply.
464
        output_parallel = self.apply(input_, bias)
465
466
467
468
469
470
471
472
473
        if self.base_layer.gather_output:
            # All-gather across the partitions.
            output = tensor_model_parallel_all_gather(output_parallel)
        else:
            output = output_parallel
        output_bias = (self.base_layer.bias
                       if self.base_layer.skip_bias_add else None)
        return output, output_bias

474
    @classmethod
475
    @_not_fully_sharded_can_replace
476
477
478
479
480
481
482
    def can_replace_layer(cls, source_layer: nn.Module,
                          lora_config: LoRAConfig, packed_modules_list: List,
                          model_config: Optional[PretrainedConfig]) -> bool:
        return type(source_layer) is ColumnParallelLinear or (
            type(source_layer) is MergedColumnParallelLinear
            and len(packed_modules_list) == 1)

483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500

class MergedColumnParallelLinearWithLoRA(ColumnParallelLinearWithLoRA):
    """ColumnParallelLinear layer that is composed of 2 sublayers (slices)
    packed together (eg. gate_proj + up_proj -> gate_up_proj).

    This means we have 2 LoRAs, each applied to one half of the layer.

    Both slices must have the same size.
    """

    def __init__(self, base_layer: MergedColumnParallelLinear) -> None:
        super().__init__(base_layer)

    def create_lora_weights(
            self,
            max_loras: int,
            lora_config: LoRAConfig,
            model_config: Optional[PretrainedConfig] = None) -> None:
501
        self.lora_config = lora_config
502
503
504
505
506
507
508
509
        n_slices = 2
        if not (len(self.base_layer.output_sizes) == n_slices
                and self.base_layer.output_sizes[0]
                == self.base_layer.output_sizes[1]):
            raise ValueError(
                "LoRAColumnParallelLinear2Slice requires 2 slices with "
                "the same size.")
        self.tp_size = get_tensor_model_parallel_world_size()
510
511
512
513
514
        self.tp_rank = get_tensor_model_parallel_rank()

        lora_a_output_size_per_partition = (
            lora_config.max_lora_rank if not lora_config.fully_sharded_loras
            else divide(lora_config.max_lora_rank, self.tp_size))
515
516
517
518
519

        self.lora_a_stacked = tuple(
            torch.zeros(
                max_loras,
                1,
520
                lora_a_output_size_per_partition,
521
                self.input_size,
522
                dtype=lora_config.lora_dtype,
523
                device=self.device,
524
525
526
527
528
            ) for _ in range(n_slices))
        self.lora_b_stacked = tuple(
            torch.zeros(
                max_loras,
                1,
529
                self.output_size // 2,
530
531
                lora_config.max_lora_rank,
                dtype=lora_config.lora_dtype,
532
                device=self.device,
533
534
535
            ) for _ in range(n_slices))

        self.output_dim = self.lora_b_stacked[0].shape[2]
536
537
        # Lazily initialized.
        self.indices: torch.Tensor
538
539
540
541
542
543
544

    def reset_lora(self, index: int):
        self.lora_a_stacked[0][index] = 0
        self.lora_a_stacked[1][index] = 0
        self.lora_b_stacked[0][index] = 0
        self.lora_b_stacked[1][index] = 0

545
546
547
    def slice_lora_a(
        self, lora_a: List[Union[torch.Tensor, None]]
    ) -> List[Union[torch.Tensor, None]]:
548
549
        return lora_a

550
551
552
553
554
    def slice_lora_b(
        self, lora_b: List[Union[torch.Tensor, None]]
    ) -> List[Union[torch.Tensor, None]]:
        if lora_b[0] is None or lora_b[1] is None:
            return lora_b
555
556
557
558
559
560
561
562
        shard_size = self.output_dim
        start_idx = self.tp_rank * shard_size
        end_idx = (self.tp_rank + 1) * shard_size
        lora_b = [
            lora_b[0][:, start_idx:end_idx], lora_b[1][:, start_idx:end_idx]
        ]
        return lora_b

563
564
565
566
567
568
569
570
571
572
    def set_lora(
        self,
        index: int,
        lora_a: torch.Tensor,
        lora_b: torch.Tensor,
        embeddings_tensor: Optional[torch.Tensor],
    ):
        self.reset_lora(index)

        if self.tp_size > 1:
573
574
            lora_a = self.slice_lora_a(lora_a)
            lora_b = self.slice_lora_b(lora_b)
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590

        if lora_a[0] is not None:
            self.lora_a_stacked[0][
                index, 0, :lora_a[0].shape[1], :lora_a[0].shape[0]].copy_(
                    lora_a[0].T, non_blocking=True)
            self.lora_b_stacked[0][
                index, 0, :lora_b[0].shape[1], :lora_b[0].shape[0]].copy_(
                    lora_b[0].T, non_blocking=True)
        if lora_a[1] is not None:
            self.lora_a_stacked[1][
                index, 0, :lora_a[1].shape[1], :lora_a[1].shape[0]].copy_(
                    lora_a[1].T, non_blocking=True)
            self.lora_b_stacked[1][
                index, 0, :lora_b[1].shape[1], :lora_b[1].shape[0]].copy_(
                    lora_b[1].T, non_blocking=True)

591
592
593
    def apply(self, x: torch.Tensor,
              bias: Optional[torch.Tensor]) -> torch.Tensor:
        output = self.base_layer.quant_method.apply(self.base_layer, x, bias)
594
595
596
597
598
599
600
601
602
603
        _apply_lora_packed_nslice(
            x,
            self.lora_a_stacked,
            self.lora_b_stacked,
            self.indices[:self.indices_len[0]],
            output,
            (self.output_dim, self.output_dim),
        )
        return output

604
    @classmethod
605
    @_not_fully_sharded_can_replace
606
607
608
609
610
611
    def can_replace_layer(cls, source_layer: nn.Module,
                          lora_config: LoRAConfig, packed_modules_list: List,
                          model_config: Optional[PretrainedConfig]) -> bool:
        return type(source_layer) is MergedColumnParallelLinear and len(
            packed_modules_list) == 2

612
613

class QKVParallelLinearWithLora(ColumnParallelLinearWithLoRA):
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
    """
    ColumnParallelLinear layer that is specifically designed for  
    qkv_proj. Certain models, such as chtglm3 and baichuan-7b,  
    only contains a single LoRA within their qkv_proj layer. 

    During inference with Tensor Parallel, the weights of lora_b 
    must be accurately partitioned according to the respective ranks.
    
    Q slice may have different shape than K and V slices (which both have
    the same shape).
    """

    def __init__(self, base_layer: QKVParallelLinear) -> None:
        super().__init__(base_layer)
        self.tp_size = get_tensor_model_parallel_world_size()
        self.q_proj_total_size = (self.base_layer.total_num_heads *
                                  self.base_layer.head_size)
        self.q_proj_shard_size = (self.base_layer.num_heads *
                                  self.base_layer.head_size)
        self.kv_proj_shard_size = (self.base_layer.num_kv_heads *
                                   self.base_layer.head_size)
        self.kv_proj_total_size = (self.base_layer.total_num_kv_heads *
                                   self.base_layer.head_size)

638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
    def slice_lora_b(self, lora_b: torch.Tensor) -> torch.Tensor:
        tp_rank = get_tensor_model_parallel_rank()
        self.q_shard_id = tp_rank
        self.kv_shard_id = tp_rank // self.base_layer.num_kv_head_replicas
        lora_b_q = lora_b[:, self.q_proj_shard_size *
                          self.q_shard_id:self.q_proj_shard_size *
                          (self.q_shard_id + 1)]
        k_offset = self.q_proj_total_size
        lora_b_k = lora_b[:, k_offset +
                          self.kv_proj_shard_size * self.kv_shard_id:k_offset +
                          self.kv_proj_shard_size * (self.kv_shard_id + 1)]
        v_offset = k_offset + self.kv_proj_total_size
        lora_b_v = lora_b[:, v_offset +
                          self.kv_proj_shard_size * self.kv_shard_id:v_offset +
                          self.kv_proj_shard_size * (self.kv_shard_id + 1)]
        lora_b = torch.cat([lora_b_q, lora_b_k, lora_b_v], dim=1)
        return lora_b

656
657
658
659
660
661
662
663
664
    def set_lora(
        self,
        index: int,
        lora_a: torch.Tensor,
        lora_b: torch.Tensor,
        embeddings_tensor: Optional[torch.Tensor],
    ):
        self.reset_lora(index)
        if self.tp_size > 1:
665
666
            lora_a = self.slice_lora_a(lora_a)
            lora_b = self.slice_lora_b(lora_b)
667
668
669
670
671
672
673
674
675

        self.lora_a_stacked[index,
                            0, :lora_a.shape[1], :lora_a.shape[0]].copy_(
                                lora_a.T, non_blocking=True)
        self.lora_b_stacked[index,
                            0, :lora_b.shape[1], :lora_b.shape[0]].copy_(
                                lora_b.T, non_blocking=True)

    @classmethod
676
    @_not_fully_sharded_can_replace
677
678
679
680
681
682
683
684
    def can_replace_layer(cls, source_layer: nn.Module,
                          lora_config: LoRAConfig, packed_modules_list: List,
                          model_config: Optional[PretrainedConfig]) -> bool:
        return type(source_layer) is QKVParallelLinear and len(
            packed_modules_list) == 1


class MergedQKVParallelLinearWithLora(ColumnParallelLinearWithLoRA):
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
    """ColumnParallelLinear layer that is composed of 3 sublayers (slices)
    packed together in qkv proj fashion
    (q_proj + k_proj + v_proj -> qkv_proj).

    This means we have 3 LoRAs, each applied to one slice of the layer.

    Q slice may have different shape than K and V slices (which both have
    the same shape).
    """

    def __init__(self, base_layer: QKVParallelLinear) -> None:
        super().__init__(base_layer)

    def create_lora_weights(
            self,
            max_loras: int,
            lora_config: LoRAConfig,
            model_config: Optional[PretrainedConfig] = None) -> None:
703
        self.lora_config = lora_config
704
        self.tp_size = get_tensor_model_parallel_world_size()
705
        self.tp_rank = get_tensor_model_parallel_rank()
706
707
708
709
        self.q_proj_shard_size = (self.base_layer.num_heads *
                                  self.base_layer.head_size)
        self.kv_proj_shard_size = (self.base_layer.num_kv_heads *
                                   self.base_layer.head_size)
710
711
        self.q_shard_id = self.tp_rank
        self.kv_shard_id = self.tp_rank // self.base_layer.num_kv_head_replicas
712

713
714
715
        lora_a_output_size_per_partition = (
            lora_config.max_lora_rank if not lora_config.fully_sharded_loras
            else divide(lora_config.max_lora_rank, self.tp_size))
716
717
718
719
720
        # q, k, v
        self.lora_a_stacked = (
            torch.zeros(
                max_loras,
                1,
721
                lora_a_output_size_per_partition,
722
                self.input_size,
723
                dtype=lora_config.lora_dtype,
724
                device=self.device,
725
726
727
728
            ),
            torch.zeros(
                max_loras,
                1,
729
                lora_a_output_size_per_partition,
730
                self.input_size,
731
                dtype=lora_config.lora_dtype,
732
                device=self.device,
733
734
735
736
            ),
            torch.zeros(
                max_loras,
                1,
737
                lora_a_output_size_per_partition,
738
                self.input_size,
739
                dtype=lora_config.lora_dtype,
740
                device=self.device,
741
742
743
744
745
746
747
748
749
            ),
        )
        self.lora_b_stacked = (
            torch.zeros(
                max_loras,
                1,
                self.q_proj_shard_size,
                lora_config.max_lora_rank,
                dtype=lora_config.lora_dtype,
750
                device=self.device,
751
752
753
754
755
756
757
            ),
            torch.zeros(
                max_loras,
                1,
                self.kv_proj_shard_size,
                lora_config.max_lora_rank,
                dtype=lora_config.lora_dtype,
758
                device=self.device,
759
760
761
762
763
764
765
            ),
            torch.zeros(
                max_loras,
                1,
                self.kv_proj_shard_size,
                lora_config.max_lora_rank,
                dtype=lora_config.lora_dtype,
766
                device=self.device,
767
768
769
770
771
772
773
            ),
        )

        self.output_slices = (self.q_proj_shard_size, self.kv_proj_shard_size,
                              self.kv_proj_shard_size)
        self.packed_indices: Optional[torch.Tensor] = None
        self.standard_indices: Optional[torch.Tensor] = None
774
775
        # lazily initialized.
        self.indices_len: List[int]
776
777
778
779
780
781
782
783
784

    def reset_lora(self, index: int):
        self.lora_a_stacked[0][index] = 0
        self.lora_b_stacked[0][index] = 0
        self.lora_a_stacked[1][index] = 0
        self.lora_b_stacked[1][index] = 0
        self.lora_a_stacked[2][index] = 0
        self.lora_b_stacked[2][index] = 0

785
786
787
    def slice_lora_a(
        self, lora_a: List[Union[torch.Tensor, None]]
    ) -> List[Union[torch.Tensor, None]]:
788
789
        return lora_a

790
791
792
793
    def slice_lora_b(
        self, lora_b: List[Union[torch.Tensor, None]]
    ) -> List[Union[torch.Tensor, None]]:
        lora_b_q, lora_b_k, lora_b_v = None, None, None
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
        if lora_b[0] is not None:
            lora_b_q = lora_b[0][:, self.q_proj_shard_size *
                                 self.q_shard_id:self.q_proj_shard_size *
                                 (self.q_shard_id + 1)]
        if lora_b[1] is not None:
            lora_b_k = lora_b[1][:, self.kv_proj_shard_size *
                                 self.kv_shard_id:self.kv_proj_shard_size *
                                 (self.kv_shard_id + 1)]
        if lora_b[2] is not None:
            lora_b_v = lora_b[2][:, self.kv_proj_shard_size *
                                 self.kv_shard_id:self.kv_proj_shard_size *
                                 (self.kv_shard_id + 1)]
        lora_b = [lora_b_q, lora_b_k, lora_b_v]
        return lora_b

809
810
811
812
813
814
815
816
817
818
    def set_lora(
        self,
        index: int,
        lora_a: torch.Tensor,
        lora_b: torch.Tensor,
        embeddings_tensor: Optional[torch.Tensor],
    ):
        self.reset_lora(index)

        if self.tp_size > 1:
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
            lora_a = self.slice_lora_a(lora_a)
            lora_b = self.slice_lora_b(lora_b)

        if lora_b[0] is not None:
            lora_b_q = lora_b[0]
            self.lora_b_stacked[0][
                index, 0, :lora_b_q.shape[1], :lora_b_q.shape[0]].copy_(
                    lora_b_q.T, non_blocking=True)
        if lora_b[1] is not None:
            lora_b_k = lora_b[1]
            self.lora_b_stacked[1][
                index, 0, :lora_b_k.shape[1], :lora_b_k.shape[0]].copy_(
                    lora_b_k.T, non_blocking=True)
        if lora_b[2] is not None:
            lora_b_v = lora_b[2]
            self.lora_b_stacked[2][
                index, 0, :lora_b_v.shape[1], :lora_b_v.shape[0]].copy_(
                    lora_b_v.T, non_blocking=True)
837
838
839
840
841
842
843
844
845
846
847
848
849
850

        if lora_a[0] is not None:
            self.lora_a_stacked[0][
                index, 0, :lora_a[0].shape[1], :lora_a[0].shape[0]].copy_(
                    lora_a[0].T, non_blocking=True)
        if lora_a[1] is not None:
            self.lora_a_stacked[1][
                index, 0, :lora_a[1].shape[1], :lora_a[1].shape[0]].copy_(
                    lora_a[1].T, non_blocking=True)
        if lora_a[2] is not None:
            self.lora_a_stacked[2][
                index, 0, :lora_a[2].shape[1], :lora_a[2].shape[0]].copy_(
                    lora_a[2].T, non_blocking=True)

851
852
853
    def apply(self, x: torch.Tensor,
              bias: Optional[torch.Tensor]) -> torch.Tensor:
        output = self.base_layer.quant_method.apply(self.base_layer, x, bias)
854
855
856
857
858
859
860
861
862
863
        _apply_lora_packed_nslice(
            x,
            self.lora_a_stacked,
            self.lora_b_stacked,
            self.indices[:self.indices_len[0]],
            output,
            self.output_slices,
        )
        return output

864
    @classmethod
865
    @_not_fully_sharded_can_replace
866
867
868
869
870
871
    def can_replace_layer(cls, source_layer: nn.Module,
                          lora_config: LoRAConfig, packed_modules_list: List,
                          model_config: Optional[PretrainedConfig]) -> bool:
        return type(source_layer) is QKVParallelLinear and len(
            packed_modules_list) == 3

872
873
874
875
876
877

class RowParallelLinearWithLoRA(BaseLayerWithLoRA):

    def __init__(self, base_layer: RowParallelLinear) -> None:
        super().__init__()
        self.base_layer = base_layer
878
879
880
        self.input_size = self.base_layer.input_size_per_partition
        self.output_size = self.base_layer.output_size
        self.device = _get_lora_device(self.base_layer)
881
882
883
884
885
886

    def create_lora_weights(
            self,
            max_loras: int,
            lora_config: LoRAConfig,
            model_config: Optional[PretrainedConfig] = None) -> None:
887
888
        self.lora_config = lora_config
        self.tp_rank = get_tensor_model_parallel_rank()
889
890
891
892
893
        self.lora_a_stacked = torch.zeros(
            (
                max_loras,
                1,
                lora_config.max_lora_rank,
894
                self.input_size,
895
896
            ),
            dtype=lora_config.lora_dtype,
897
            device=self.device,
898
        )
899
900
901
902
903
        tp_size = get_tensor_model_parallel_world_size()
        lora_b_output_size_per_partition = (
            self.output_size if not lora_config.fully_sharded_loras else
            divide(self.output_size, tp_size))

904
905
906
907
        self.lora_b_stacked = torch.zeros(
            (
                max_loras,
                1,
908
                lora_b_output_size_per_partition,
909
910
911
                lora_config.max_lora_rank,
            ),
            dtype=lora_config.lora_dtype,
912
            device=self.device,
913
        )
914
915
916
        # Lazily initialized
        self.indices: torch.Tensor
        self.indices_len: List[int]
917
918
919
920
921

    def reset_lora(self, index: int):
        self.lora_a_stacked[index] = 0
        self.lora_b_stacked[index] = 0

922
923
924
925
926
927
928
929
930
931
932
    def slice_lora_a(self, lora_a: torch.Tensor) -> torch.Tensor:
        tensor_model_parallel_rank = get_tensor_model_parallel_rank()
        shard_size = self.input_size
        start_idx = tensor_model_parallel_rank * shard_size
        end_idx = (tensor_model_parallel_rank + 1) * shard_size
        lora_a = lora_a[start_idx:end_idx, :]
        return lora_a

    def slice_lora_b(self, lora_b: torch.Tensor) -> torch.Tensor:
        return lora_b

933
934
935
936
937
938
939
940
    def set_lora(
        self,
        index: int,
        lora_a: torch.Tensor,
        lora_b: torch.Tensor,
        embeddings_tensor: Optional[torch.Tensor],
    ):
        self.reset_lora(index)
941

942
        if self.base_layer.tp_size > 1:
943
944
            lora_a = self.slice_lora_a(lora_a)
            lora_b = self.slice_lora_b(lora_b)
945
946
947
948
949
950
951
952
953
954
955
956
957
958

        self.lora_a_stacked[index,
                            0, :lora_a.shape[1], :lora_a.shape[0]].copy_(
                                lora_a.T, non_blocking=True)
        self.lora_b_stacked[index,
                            0, :lora_b.shape[1], :lora_b.shape[0]].copy_(
                                lora_b.T, non_blocking=True)

    def set_mapping(
        self,
        base_indices: torch.Tensor,
        sampler_indices: torch.Tensor,
        sampler_indices_padded: torch.Tensor,
        embeddings_indices: torch.Tensor,
959
        long_lora_indices: torch.Tensor,
960
961
962
963
964
        indices_len: List[int],
    ):
        self.indices = base_indices
        self.indices_len = indices_len

965
966
    def apply(self, x: torch.Tensor) -> torch.Tensor:
        output = self.base_layer.quant_method.apply(self.base_layer, x)
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
        _apply_lora(
            x,
            self.lora_a_stacked,
            self.lora_b_stacked,
            self.indices[:self.indices_len[0]],
            output,
        )
        return output

    def forward(self, input_):
        """Forward of RowParallelLinear

        Args:
            input_: tensor whose last dimension is `input_size`. If
                    `input_is_parallel` is set, then the last dimension
                    is `input_size // tp_size`.

        Returns:
            - output
            - bias
        """
        # Set up backprop all-reduce.
        if self.base_layer.input_is_parallel:
            input_parallel = input_
        else:
            # TODO: simplify code below
            tp_rank = get_tensor_model_parallel_rank()
            splitted_input = split_tensor_along_last_dim(
                input_, num_partitions=self.base_layer.tp_size)
            input_parallel = splitted_input[tp_rank].contiguous()

        # Matrix multiply.
999
        output_parallel = self.apply(input_parallel)
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
        if self.base_layer.reduce_results and self.base_layer.tp_size > 1:
            output_ = tensor_model_parallel_all_reduce(output_parallel)
        else:
            output_ = output_parallel

        if not self.base_layer.skip_bias_add:
            output = (output_ + self.base_layer.bias
                      if self.base_layer.bias is not None else output_)
            output_bias = None
        else:
            output = output_
            output_bias = self.base_layer.bias
        return output, output_bias

    @property
    def weight(self):
1016
1017
        return self.base_layer.weight if hasattr(
            self.base_layer, "weight") else self.base_layer.qweight
1018

1019
    @classmethod
1020
    @_not_fully_sharded_can_replace
1021
1022
1023
1024
1025
    def can_replace_layer(cls, source_layer: nn.Module,
                          lora_config: LoRAConfig, packed_modules_list: List,
                          model_config: Optional[PretrainedConfig]) -> bool:
        return type(source_layer) is RowParallelLinear

1026

1027
class LogitsProcessorWithLoRA(BaseLayerWithLoRA):
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
    """
    LoRA wrapper for LogitsProcessor, with extra logic to handle the
    application of the LoRA adapter and added LoRA vocabulary.

    Args:
        base_layer: LogitsProcessor layer
        hidden_size: hidden size of the model
        dtype: data type of the model
        device: device of the model
        sharded_to_full_mapping: index mapping from sharded vocab to full vocab
            received from base_layer.get_sharded_to_full_mapping(). If None,
            no reindexing will be done.
    """
1041

1042
1043
1044
    def __init__(self, base_layer: LogitsProcessor, hidden_size: int,
                 dtype: torch.dtype, device: torch.device,
                 sharded_to_full_mapping: Optional[List[int]]) -> None:
1045
1046
1047
1048
1049
        super().__init__()
        self.base_layer = base_layer
        self.hidden_size = hidden_size
        self.dtype = dtype
        self.device = device
1050
1051
1052
        self.tp_size = get_tensor_model_parallel_world_size()
        self.tp_rank = get_tensor_model_parallel_rank()
        self.sharded_to_full_mapping = sharded_to_full_mapping
1053

1054
    @property
1055
1056
    def logits_as_input(self):
        return self.base_layer.logits_as_input
1057

1058
1059
1060
1061
    @property
    def vocab_size(self):
        return self.base_layer.vocab_size

1062
1063
1064
1065
    @property
    def scale(self):
        return self.base_layer.scale

Woosuk Kwon's avatar
Woosuk Kwon committed
1066
1067
1068
1069
    @property
    def soft_cap(self):
        return self.base_layer.soft_cap

1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
    @property
    def org_vocab_size(self):
        return self.base_layer.org_vocab_size

    @property
    def include_gpu_probs_tensor(self):
        return self.base_layer.include_gpu_probs_tensor

    def create_lora_weights(
        self,
        max_loras: int,
        lora_config: LoRAConfig,
        model_config: Optional[PretrainedConfig] = None,
    ) -> None:
        # Keep this in sync with csrc/punica/bgmv/bgmv_config.h
1085
        if 32000 < self.base_layer.vocab_size > 128512:
1086
            raise ValueError("When using LoRA, vocab size must be "
1087
                             "32000 >= vocab_size <= 128512")
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
        self.lora_a_stacked = torch.zeros(
            (
                max_loras,
                1,
                lora_config.max_lora_rank,
                self.hidden_size,
            ),
            dtype=lora_config.lora_dtype,
            device=self.device,
        )
        self.lora_b_stacked = torch.zeros(
            (
                max_loras,
                1,
                # Pad for kernel compatibility
                math.ceil(self.base_layer.vocab_size /
                          lora_config.lora_vocab_padding_size) *
                lora_config.lora_vocab_padding_size,
                lora_config.max_lora_rank,
            ),
            dtype=lora_config.lora_dtype,
            device=self.device,
        )
        self.embeddings_tensors = torch.full(
            (max_loras, lora_config.lora_extra_vocab_size, self.hidden_size),
            fill_value=float("-inf"),
            dtype=self.dtype,
            device=self.device,
        )
1117
1118
1119
1120
1121
1122
1123
        if self.sharded_to_full_mapping is not None:
            self.sharded_to_full_mapping_gpu = torch.tensor(
                self.sharded_to_full_mapping,
                device=self.device,
                dtype=torch.long)
        else:
            self.sharded_to_full_mapping_gpu = None
1124
1125
1126
1127
        # Lazily initialized.
        self.indices: torch.Tensor
        self.indices_len: List[int]
        self.indices_padded: torch.Tensor
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

    def reset_lora(self, index: int):
        self.lora_a_stacked[index] = 0
        self.lora_b_stacked[index] = 0
        self.embeddings_tensors[index] = float("-inf")

    def set_lora(
        self,
        index: int,
        lora_a: torch.Tensor,
        lora_b: torch.Tensor,
        embeddings_tensor: Optional[torch.Tensor],
    ):
        self.reset_lora(index)
        self.lora_a_stacked[index,
                            0, :lora_a.shape[1], :lora_a.shape[0]].copy_(
                                lora_a.T, non_blocking=True)
        self.lora_b_stacked[index,
                            0, :lora_b.shape[1], :lora_b.shape[0]].copy_(
                                lora_b.T, non_blocking=True)
        if embeddings_tensor is not None:
            self.embeddings_tensors[
                index, :embeddings_tensor.shape[0], :embeddings_tensor.
                shape[1], ] = embeddings_tensor

    def set_mapping(
        self,
        base_indices: torch.Tensor,
        sampler_indices: torch.Tensor,
        sampler_indices_padded: torch.Tensor,
        embeddings_indices: torch.Tensor,
1159
        long_lora_indices: torch.Tensor,
1160
1161
1162
1163
1164
1165
1166
1167
1168
        indices_len: List[int],
    ):
        self.indices = sampler_indices
        self.indices_padded = sampler_indices_padded
        self.indices_len = indices_len

    def _get_logits(
        self,
        hidden_states: torch.Tensor,
1169
        lm_head: VocabParallelEmbedding,
1170
        embedding_bias: Optional[torch.Tensor] = None,
1171
    ) -> Optional[torch.Tensor]:
1172
        # Get the logits for the next tokens.
1173
        logits = lm_head.linear_method.apply(lm_head, hidden_states)
1174
1175
1176
1177
1178
1179
        if embedding_bias is not None:
            logits += embedding_bias
        logits = tensor_model_parallel_gather(logits)
        if logits is None:
            return None

1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
        if self.sharded_to_full_mapping_gpu is not None:
            # Reindex full logits tensor to ensure 1:1 mapping between
            # index and token_id
            # Example for:
            #   org_vocab_size = 4
            #   added_vocab_size = 2
            #   pad_to_size = 8
            #   tp_size = 2

            # indices:  [0, 1, 2,  3, 4, 5, 6,  7]
            # token_id: [0, 1, 4, -1, 2, 3, 5, -1]

            # Therefore, the mapping is expected to be:
            # [0, 1, 4, 6, 2, 3, 5, 7] so that when we reindex,
            # we get:
            # indices:  [0, 1, 2, 3, 4, 5,  6,  7]
            # token_id: [0, 1, 2, 3, 4, 5, -1, -1]
            logits = logits[:, self.sharded_to_full_mapping_gpu]

1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
        lora_logits = torch.empty(
            self.embeddings_tensors.shape[0] + 1,
            self.embeddings_tensors.shape[1],
            hidden_states.shape[0],
            dtype=self.embeddings_tensors.dtype,
            device=self.embeddings_tensors.device,
        )
        torch.matmul(self.embeddings_tensors,
                     hidden_states.T,
                     out=lora_logits[:-1])
        lora_logits[-1] = float("-inf")
        lora_logits = lora_logits.mT
        lora_logits = (lora_logits.reshape(
            lora_logits.shape[0] * lora_logits.shape[1],
            lora_logits.shape[2],
        ).index_select(0,
                       self.indices_padded[:self.indices_len[2]]).nan_to_num_(
                           nan=float("-inf"),
                           posinf=float("inf"),
                           neginf=float("-inf")))
        logits[:,
               self.base_layer.org_vocab_size:self.base_layer.org_vocab_size +
               lora_logits.shape[1]] = lora_logits

        _apply_lora(
            hidden_states,
            self.lora_a_stacked,
            self.lora_b_stacked,
            self.indices[:self.indices_len[1]],
            logits,
        )

        # Remove paddings in vocab (if any).
        logits = logits[:, :self.base_layer.vocab_size]

        return logits

    def forward(self, *args, **kwargs):
        return type(self.base_layer).forward(self, *args, **kwargs)

1239
1240
1241
1242
1243
1244
    @classmethod
    def can_replace_layer(cls, source_layer: nn.Module,
                          lora_config: LoRAConfig, packed_modules_list: List,
                          model_config: Optional[PretrainedConfig]) -> bool:
        # Special handling for the LogitsProcessor.
        return False
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342


class LinearScalingRotaryEmbeddingWithLora(BaseLayerWithLoRA):
    """Implements RoPE-scaled embeddings with linear scaling for
    multiple LoRA adapters with a specialized kernel.

    Replace LinearScalingRotaryEmbedding with MultiLinearScalingRotaryEmbedding
    which can handle multi lora adapters in a specialied kernel.
    """

    def __init__(self, base_layer: RotaryEmbedding) -> None:
        super().__init__()
        self.base_layer = base_layer
        # Lazily initialized
        self.long_lora_indices: torch.Tensor
        self.indices_len: List[int]

    @property
    def scaling_factors(self):
        return self.base_layer.scaling_factors

    @property
    def rotary_dim(self):
        return self.base_layer.rotary_dim

    def create_lora_weights(
        self,
        max_loras: int,
        lora_config: LoRAConfig,
        model_config: Optional[PretrainedConfig] = None,
    ) -> None:
        scaling_factors = list(
            lora_config.long_lora_scaling_factors
        ) if lora_config.long_lora_scaling_factors else []
        base_scaling_factor = (self.base_layer.scaling_factor if isinstance(
            self.base_layer, LinearScalingRotaryEmbedding) else 1.0)
        scaling_factors = sorted(
            list(set([base_scaling_factor] + scaling_factors)))
        self.base_layer = LinearScalingRotaryEmbedding(
            self.base_layer.head_size,
            self.base_layer.rotary_dim,
            self.base_layer.max_position_embeddings,
            self.base_layer.base,
            self.base_layer.is_neox_style,
            scaling_factors,
            self.base_layer.dtype,
        )

    def reset_lora(self, index: int):
        ...

    def set_lora(
        self,
        index: int,
        lora_a: torch.Tensor,
        lora_b: torch.Tensor,
        embeddings_tensor: Optional[torch.Tensor],
    ):
        ...

    def set_mapping(
        self,
        base_indices: torch.Tensor,
        sampler_indices: torch.Tensor,
        sampler_indices_padded: torch.Tensor,
        embeddings_indices: torch.Tensor,
        long_lora_indices: torch.Tensor,
        indices_len: List[int],
    ):
        self.long_lora_indices = long_lora_indices
        self.indices_len = indices_len

    def forward(
        self,
        positions: torch.Tensor,
        query: torch.Tensor,
        key: torch.Tensor,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        return self.base_layer(
            positions,
            query,
            key,
            offsets=self.long_lora_indices[:self.indices_len[4]])

    @property
    def scaling_factor_to_offset(self) -> Dict[float, int]:
        return self.base_layer.scaling_factor_to_offset

    @classmethod
    def can_replace_layer(cls, source_layer: nn.Module,
                          lora_config: LoRAConfig, packed_modules_list: List,
                          model_config: Optional[PretrainedConfig]) -> bool:
        """Returns True if the layer can be replaced by this LoRA layer."""
        return type(source_layer) is LinearScalingRotaryEmbedding or type(
            source_layer) is RotaryEmbedding

    def extra_repr(self) -> str:
        return self.base_layer.extra_repr()