model_manager.py 33.3 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import math
5
6
from collections.abc import Callable
from typing import TypeVar
7

8
import regex as re
9
10
11
import torch
from torch import nn

12
13
from vllm.config import VllmConfig
from vllm.config.lora import LoRAConfig, ModelConfig
14
from vllm.logger import init_logger
15
16
17
18
19
20
from vllm.lora.layers import (
    BaseLayerWithLoRA,
    FusedMoE3DWithLoRA,
    LoRAMapping,
    LoRAMappingType,
)
21
from vllm.lora.lora_model import LoRAModel
22
from vllm.lora.lora_weights import LoRALayerWeights, PackedLoRALayerWeights
23
from vllm.lora.punica_wrapper import PunicaWrapperBase, get_punica_wrapper
24
25
26
27
from vllm.lora.utils import (
    from_layer,
    from_layer_logits_processor,
    get_supported_lora_modules,
28
    is_moe_model,
29
    process_packed_modules_mapping,
30
31
    replace_submodule,
)
32
from vllm.model_executor.layers.fused_moe import FusedMoE
33
from vllm.model_executor.models import SupportsLoRA, supports_multimodal
34
from vllm.model_executor.models.interfaces import is_pooling_model
35
from vllm.model_executor.models.module_mapping import MultiModelKeys
36
from vllm.model_executor.models.utils import PPMissingLayer
37
from vllm.multimodal import MULTIMODAL_REGISTRY
38
from vllm.utils.cache import LRUCache
39
from vllm.utils.platform_utils import is_pin_memory_available
40
from vllm.v1.worker.utils import MultiModalBudget
41

42
logger = init_logger(__name__)
43

44
T = TypeVar("T")
45
DEFAULT_LANGUAGE_WRAPPER_KEY = "language_model"
46
47
48
49
50
51
52


class AdapterLRUCache(LRUCache[int, T]):
    def __init__(self, capacity: int, deactivate_fn: Callable[[int], object]):
        super().__init__(capacity)
        self.deactivate_fn = deactivate_fn

53
    def _on_remove(self, key: int, value: T | None):
54
55
56
57
58
59
        logger.debug("Removing adapter int id: %d", key)
        self.deactivate_fn(key)
        return super()._on_remove(key, value)


class LoRAModelManager:
60
61
62
63
    """A manager that manages multiple LoRA-fine-tuned models."""

    def __init__(
        self,
64
        model: SupportsLoRA,
65
66
67
68
        max_num_seqs: int,
        max_num_batched_tokens: int,
        vocab_size: int,
        lora_config: LoRAConfig,
69
        device: torch.device,
70
        vllm_config: VllmConfig | None = None,
71
72
73
74
75
76
77
78
79
80
81
82
    ):
        """Create a LoRAModelManager and adapter for a given model.

        Args:
            model: the model to be adapted.
            max_num_seqs: the maximum number of sequences model can run in a
                single batch.
            max_num_batched_tokens: the maximum number of tokens model can run
                in a single batch.
            vocab_size: the vocab size of the model.
            lora_config: the LoRA configuration.
        """
83
        self.model: SupportsLoRA = model
84
85
86
87
88
        self.supported_lora_modules = get_supported_lora_modules(self.model)
        assert self.supported_lora_modules, (
            f"No supported LoRA modules found in {self.model.__class__.__name__}."
        )

89
90
91
92
        self._registered_adapters: dict[int, LoRAModel] = {}
        # Dict instead of a set for compatibility with LRUCache.
        self._active_adapters: dict[int, None] = {}
        self.adapter_type = "LoRA"
93
        self.lora_config = lora_config
94
        self.device = device
95
96
97
        self.max_num_seqs = max_num_seqs
        assert self.capacity >= self.lora_slots
        self.max_num_batched_tokens = math.ceil(max_num_batched_tokens / 8) * 8
98
        self.lora_index_to_id: list[int | None] = [None] * self.lora_slots
99
        self.vocab_size = vocab_size
100
        self.packed_modules_mapping = process_packed_modules_mapping(self.model)
101

102
        self.is_pooling_model = is_pooling_model(self.model)
103
104
105
        self.packed_modules: dict[str, list[str]] = {}
        self.modules: dict[str, BaseLayerWithLoRA] = {}
        # Dict instead of a set for compatibility with LRUCache.
106
        self._last_mapping: LoRAMapping | None = None
107
        self._is_3d_moe_model = is_moe_model(self.model) and self.model.is_3d_moe_weight
108
        self._init_punica_wrapper(max_num_batched_tokens, vllm_config)
109
        self._create_lora_modules()
110

111
        self.model.lora_manager = self
112

113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
    def _init_punica_wrapper(
        self, max_num_batched_tokens: int, vllm_config: VllmConfig
    ) -> None:
        # Used to indicate whether the model is a multimodal model
        self.supports_mm: bool = (
            supports_multimodal(self.model)
            # In case the model only supports LoRA for
            # text modules (e.g. ChatGLM)
            and hasattr(self.model, "get_mm_mapping")
        )
        self.punica_wrapper_mapping: dict[str, PunicaWrapperBase] = {}
        if self.supports_mm:
            self._maybe_init_mm(vllm_config, max_num_batched_tokens)
        else:
            llm_punica_wrapper = get_punica_wrapper(
                max_num_batched_tokens,
                max_batches=self.max_num_seqs,
                device=self.device,
131
                lora_config=self.lora_config,
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
            )

            self.punica_wrapper_mapping[DEFAULT_LANGUAGE_WRAPPER_KEY] = (
                llm_punica_wrapper
            )

    def _maybe_init_mm(self, vllm_config: VllmConfig, max_num_batched_tokens) -> None:
        self.supports_tower_connector_lora = False
        model_config: ModelConfig = vllm_config.model_config
        self.mm_mapping: MultiModelKeys = self.model.get_mm_mapping()

        # Only one language model can be included in the model.
        assert len(self.mm_mapping.language_model) == 1

        # Language model punica wrapper
        llm_punica_wrapper = get_punica_wrapper(
            max_num_batched_tokens,
            max_batches=self.max_num_seqs,
            device=self.device,
151
            lora_config=self.lora_config,
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
        )
        lm_prefix = self.mm_mapping.language_model[0]
        self.punica_wrapper_mapping[lm_prefix] = llm_punica_wrapper

        if self.lora_config.enable_tower_connector_lora:
            self.mm_processor_info = MULTIMODAL_REGISTRY.create_processor(
                model_config
            ).info
            self.supports_tower_connector_lora = self.supports_mm and hasattr(
                self.model, "get_num_mm_encoder_tokens"
            )
        if not self.supports_tower_connector_lora:
            return

        logger.warning(
            "LoRA for the tower and connector of multimodal models is "
            "experimental and may contain bugs. Please report any related issues on "
            "GitHub if you encounter them."
        )

        mm_budget = MultiModalBudget(
            model_config,
            vllm_config.scheduler_config,
            MULTIMODAL_REGISTRY,
        )
        limit_per_prompt: int = max(
            self.mm_processor_info.get_allowed_mm_limits().values()
        )
        num_encoder_tokens = self.model.get_num_mm_encoder_tokens(
            mm_budget.get_encoder_budget()
        )

        # Tower wrappers
        tower_punica_wrapper = get_punica_wrapper(
            num_encoder_tokens,
            max_batches=self.max_num_seqs * limit_per_prompt,
            device=self.device,
189
            lora_config=self.lora_config,
190
191
192
193
194
195
196
197
198
199
200
201
202
203
        )
        for prefix in self.mm_mapping.tower_model:
            self.punica_wrapper_mapping[prefix] = tower_punica_wrapper

        # Use wrapper for connector if present.
        if self.mm_mapping.connector:
            if hasattr(self.model, "get_num_mm_connector_tokens"):
                connector_tokens = self.model.get_num_mm_connector_tokens(
                    num_encoder_tokens
                )
                connector_punica_wrapper = get_punica_wrapper(
                    connector_tokens,
                    max_batches=self.max_num_seqs * limit_per_prompt,
                    device=self.device,
204
                    lora_config=self.lora_config,
205
206
207
208
209
210
211
212
213
214
                )
                for prefix in self.mm_mapping.connector:
                    self.punica_wrapper_mapping[prefix] = connector_punica_wrapper
            else:
                logger.warning_once(
                    "Connector LoRA support disabled: model does not implement "
                    "get_num_mm_connector_tokens(). This method is required to "
                    "determine the connector's token budget for LoRA operations."
                )

215
216
    def __len__(self) -> int:
        return len(self._registered_adapters)
217
218
219
220
221
222
223
224
225

    @property
    def capacity(self) -> int:
        return self.lora_config.max_cpu_loras

    @property
    def lora_slots(self) -> int:
        return self.lora_config.max_loras

226
227
228
    @property
    def adapter_slots(self) -> int:
        return self.lora_slots
229

230
    def activate_adapter(
231
232
233
234
        self,
        lora_id: int,
    ) -> bool:
        """Move LoRA into a GPU buffer to be used in the forward pass."""
235
        if lora_id in self._active_adapters:
236
237
            return False
        first_free_slot = next(
238
239
240
241
242
243
244
            (
                (i, lora_id)
                for i, lora_id in enumerate(self.lora_index_to_id)
                if lora_id is None
            ),
            None,
        )
245
246
247
        if first_free_slot is None:
            raise ValueError("No free lora slots")
        index, _ = first_free_slot
248
249
        self._active_adapters[lora_id] = None
        lora_model = self._registered_adapters[lora_id]
250
251
252
        logger.debug(
            "Activating LoRA. int id: %d, slot index: %d", lora_model.id, index
        )
253
254
        self.lora_index_to_id[index] = lora_model.id
        for module_name, module in self.modules.items():
255
            module_lora = self._get_lora_layer_weights(lora_model, module_name)
256
257
258
            if not module_lora:
                module.reset_lora(index)
                continue
259

260
261
262
263
264
            module.set_lora(
                index,
                module_lora.lora_a,
                module_lora.lora_b,
            )
265

266
267
        return True

268
    def _deactivate_adapter(self, lora_id: int):
269
270
271
272
273
274
        try:
            index = self.lora_index_to_id.index(lora_id)
            self.lora_index_to_id[index] = None
        except ValueError:
            pass

275
    def _add_adapter(self, lora: LoRAModel):
276
        self._create_merged_loras_inplace(lora)
277
        self._registered_adapters[lora.id] = lora
278

279
    def pin_adapter(self, lora_id: int) -> bool:
280
281
        """Pin a LoRAModel in the manager cache."""
        raise NotImplementedError(
282
            "Pinning is not supported in LoRAModelManager. "
283
284
            "Use LRUCacheLoRAModelManager for pinning"
        )  # type: ignore
285

286
    def _set_adapter_mapping(self, mapping: LoRAMapping) -> None:
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
        # Default to the main language model wrapper
        if not (self.supports_mm and self.supports_tower_connector_lora):
            target_prefix = (
                self.mm_mapping.language_model[0]
                if self.supports_mm
                else DEFAULT_LANGUAGE_WRAPPER_KEY
            )
        elif mapping.type == LoRAMappingType.TOWER and self.mm_mapping.tower_model:
            target_prefix = self.mm_mapping.tower_model[0]
        elif mapping.type == LoRAMappingType.CONNECTOR and self.mm_mapping.connector:
            target_prefix = self.mm_mapping.connector[0]
        else:
            target_prefix = self.mm_mapping.language_model[0]

        punica_wrapper = self._get_punica_wrapper(target_prefix)
        assert punica_wrapper is not None

        punica_wrapper.update_metadata(
305
306
307
308
309
            mapping,
            self.lora_index_to_id,
            self.lora_slots + 1,
            self.vocab_size,
        )
310

311
    def remove_all_adapters(self):
312
        """Remove all LoRAModels from the manager."""
313
        self._registered_adapters.clear()
314
        self.lora_index_to_id = [None] * self.lora_slots
315
        self._active_adapters.clear()
316
317

    def _create_lora_modules(self):
318
319
320
321
322
        def _parent_module(module_name: str) -> str:
            # module name is a dot separated name.
            # for example:
            #  - given an input 'x.y.z' return 'x.y'
            #  - given an input 'x' return ''
323
            return module_name.rpartition(".")[0]
324

325
        for module_name, module in self.model.named_modules(remove_duplicate=False):
326
327
            if isinstance(module, PPMissingLayer):
                continue
328

329
330
            if not self._match_target_modules(module_name):
                continue
331
332
333

            punica_wrapper = self._get_punica_wrapper(module_name)
            if punica_wrapper is None:
334
                logger.warning(
335
336
337
                    "Regarding %s, vLLM currently only supports adding LoRA to"
                    " language model, %s will be ignored.",
                    self.model.__class__.__name__,
338
339
340
                    module_name,
                )
                continue
341

342
343
            parts = module_name.split(".")[-1]
            packed_moduled_lst = self.packed_modules_mapping.get(parts, [])
344
345
346
347
348
349
350
            if isinstance(module, FusedMoE):
                # packed_moduled_lst is used here to just determine whether to
                # instantiate FusedMoE3DWithLoRA or FusedMoEWithLoRA, and the
                # difference between these two LoRA layers is whether the
                # LoRA weights of w1 and w3 have already been fused on disk.

                packed_moduled_lst = ["w13"] if self._is_3d_moe_model else ["w1", "w3"]
351
            new_module = replace_submodule(
352
353
354
355
356
357
358
359
360
361
                self.model,
                module_name,
                from_layer(
                    module,
                    self.lora_slots,
                    self.lora_config,
                    packed_moduled_lst,
                    self.model.config,
                ),
            )
362

363
364
            # (yard1): TODO make this more robust
            if "lm_head" in module_name:
365
                logits_processor_module_name = "logits_processor"
366
367
368
                parent_module = _parent_module(module_name)
                if parent_module:
                    logits_processor_module_name = (
369
370
                        f"{parent_module}.{logits_processor_module_name}"
                    )
371

372
                logits_processor_module = self.model.get_submodule(
373
374
                    logits_processor_module_name
                )
375

376
                new_module = replace_submodule(
377
378
379
380
381
382
383
384
385
386
                    self.model,
                    logits_processor_module_name,
                    from_layer_logits_processor(
                        logits_processor_module,
                        module,
                        self.lora_slots,
                        self.lora_config,
                        self.model.config,
                    ),
                )
387
388
389
390
391
392

            # In some models, especially multimodal ones, layers with the same
            # name may have different types, such as nn.Linear and
            # ReplicatedLinear. The nn.Linear layers cannot be replaced with
            # LoRA layers, leading to assertion error. The following check
            # aims to prevent this error
393
            if self.supports_mm and not isinstance(new_module, BaseLayerWithLoRA):
394
                continue
395
            self.register_module(module_name, new_module)
396

397
            self._register_packed_modules(module_name)
398
            # All lora layers share the same punica_wrapper based on reference.
399
            new_module.set_mapping(punica_wrapper)
400
401

    def register_module(self, module_name: str, module: "BaseLayerWithLoRA"):
402
        assert isinstance(module, BaseLayerWithLoRA), (
403
404
            f"Module {module_name} must be a BaseLayerWithLoRA instance, "
            f"got {type(module)}"
405
        )
406
407
        self.modules[module_name] = module

Terry's avatar
Terry committed
408
    def create_dummy_lora(
409
410
411
        self,
        lora_id: int,
        rank: int,
412
        embedding_modules: dict[str, str] | None = None,
413
    ) -> LoRAModel:
414
        """Create zero-initialized LoRAModel for warmup."""
415
        model = LoRAModel(lora_id, rank, {})
416
        for module_name, module in self.model.named_modules():
417
418
419
            if (
                not self._match_target_modules(module_name)
                or not isinstance(module, BaseLayerWithLoRA)
420
                or self._get_punica_wrapper(module_name) is None
421
            ):
422
423
424
                continue
            parts = module_name.split(".")
            if module_name not in self.packed_modules:
425
                assert embedding_modules is not None
Terry's avatar
Terry committed
426
                if parts[-1] in embedding_modules:
427
428
429
430
431
432
433
434
435
436
                    input_dim = (
                        module.base_layer.org_vocab_size
                        if hasattr(module.base_layer, "org_vocab_size")
                        else module.base_layer.weight.shape[1]
                    )
                    output_dim = (
                        module.base_layer.embedding_dim
                        if hasattr(module.base_layer, "embedding_dim")
                        else module.base_layer.weight.shape[0]
                    )
437
438
439
440
441
                    lora = LoRALayerWeights.create_dummy_lora_weights(
                        module_name,
                        input_dim,
                        output_dim,
                        rank,
442
                        module.lora_a_stacked[0].dtype,
443
                        "cpu",
444
                    )
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
                    model.loras[module_name] = lora
                elif module.__class__.__name__ == "FusedMoE3DWithLoRA":
                    # Case for 3D moe model
                    # w2
                    lora = LoRALayerWeights.create_dummy_lora_weights(
                        module_name,
                        module.w2_input_size,
                        module.w2_output_size,
                        rank * module.w2_lora_a_stacked[0].shape[1],  # rank*num_experts
                        module.w2_lora_a_stacked[0].dtype,
                        "cpu",
                    )
                    model.loras[module_name] = lora
                    # w13
                    lora = LoRALayerWeights.create_dummy_lora_weights(
                        module_name,
                        module.w13_input_size,
                        module.w13_output_size,
                        rank
                        * module.w13_lora_a_stacked[0].shape[1],  # rank*num_experts
                        module.w13_lora_a_stacked[0].dtype,
                        "cpu",
                    )
                    model.loras[module_name + ".base_layer"] = lora
469
470
471
                else:
                    lora = LoRALayerWeights.create_dummy_lora_weights(
                        module_name,
472
473
                        module.lora_a_stacked[0].shape[-1],
                        module.lora_b_stacked[0].shape[-2],
474
                        rank,
475
                        module.lora_a_stacked[0].dtype,
476
477
                        "cpu",
                    )
478
                    model.loras[module_name] = lora
479
480
481
            else:
                parts = module_name.split(".")
                replacements = self.packed_modules_mapping[parts[-1]]
482
                subloras: list[LoRALayerWeights | None] = []
483
484
485
486
487
488
489
490
491
492
                for i, r in enumerate(replacements):
                    lora = LoRALayerWeights.create_dummy_lora_weights(
                        module_name + "." + r,
                        module.lora_a_stacked[i].shape[-1],
                        module.lora_b_stacked[i].shape[-2],
                        rank,
                        module.lora_a_stacked[i].dtype,
                        "cpu",
                    )
                    subloras.append(lora)
493
494
495
496
                if module.__class__.__name__ == "FusedMoEWithLoRA":
                    lora = PackedLoRALayerWeights.pack_moe(subloras, module_name)
                else:
                    lora = PackedLoRALayerWeights.pack(subloras)
497
                model.loras[module_name] = lora
498
499
500
501
502
        return model

    def _match_target_modules(self, module_name: str):
        return any(
            re.match(
503
504
505
506
507
                r".*\.{target_module}$".format(target_module=target_module), module_name
            )
            or target_module == module_name
            for target_module in self.supported_lora_modules
        )
508

509
    def _get_punica_wrapper(self, module_name: str) -> PunicaWrapperBase | None:
510
        """
511
        Determine whether this module supports LoRA and which wrapper to use.
512
        """
513
514
515
516
517
518
519
520
521
522
523
524
        # For language model (early return)
        if not self.supports_mm:
            return self.punica_wrapper_mapping[DEFAULT_LANGUAGE_WRAPPER_KEY]

        # For multimodal model
        # NOTE Sort by prefix length (descending) to match the longest prefix first
        # e.g., 'visual.merger' should match 'visual.merger' instead of 'visual.'
        for prefix in sorted(self.punica_wrapper_mapping.keys(), key=len, reverse=True):
            if module_name.startswith(prefix):
                return self.punica_wrapper_mapping[prefix]

        return None
525

526
527
528
    def _register_packed_modules(self, module_full_name: str) -> None:
        parts = module_full_name.split(".")
        module_name = parts[-1]
529
530
531
532
        replacements = self.packed_modules_mapping.get(module_name, [])
        # When replacements is less than or equal to 1, it indicates that this
        # module is not a packed module.
        if len(replacements) <= 1:
533
534
535
536
537
538
539
540
            return
        prefix = ".".join(parts[:-1])
        self.packed_modules[module_full_name] = [
            prefix + "." + r if prefix else r for r in replacements
        ]

    def _create_merged_loras_inplace(self, lora_model: LoRAModel) -> None:
        for module_name, new_module_names in self.packed_modules.items():
541
            replacement_loras: list[LoRALayerWeights | None] = []
542
            replaced_module: set[str] = set()
543
544
            has_replacement = False
            for r in new_module_names:
545
                lora = self._get_lora_layer_weights(lora_model, r)
546
547
548
                replacement_loras.append(lora)
                if lora:
                    has_replacement = True
549
                    replaced_module.add(r)
550
551
552
553
554
555
            if not has_replacement:
                continue
            for i in range(len(replacement_loras)):
                if replacement_loras[i]:
                    continue
                replacement_loras[i] = None
556
            # HACK Temporary solution for the pool model.
557
            if self.is_pooling_model and not lora_model.check_lora_name(module_name):
558
559
560
                replaced_module_name = module_name.replace("model.", "")
                if lora_model.check_lora_name(module_name):
                    module_name = replaced_module_name
561
562
563
564
565
566
567
568
            if module_name.endswith(".experts"):
                lora_model.loras[module_name] = PackedLoRALayerWeights.pack_moe(
                    replacement_loras, module_name
                )
            else:
                lora_model.loras[module_name] = PackedLoRALayerWeights.pack(
                    replacement_loras
                )
569
570
571
            # Remove the modules that have been replaced.
            for module in replaced_module:
                lora_model.loras.pop(module, None)
572

573
574
575
        for lora in lora_model.loras.values():
            lora.optimize()

576
577
578
579
        for module_name, module in self.modules.items():
            if isinstance(module, FusedMoE3DWithLoRA):
                self._stack_moe_lora_weights(lora_model, module, module_name)

580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
        first_lora: LoRALayerWeights = next(iter(lora_model.loras.values()))
        assert first_lora.lora_a is not None
        if isinstance(first_lora.lora_a, list):
            lora_device = next(iter(first_lora.lora_a))
        else:
            lora_device = first_lora.lora_a.device
        # Execute pin_memory after LoRA weight merging, mainly because:
        # 1. Some MoE models have a large number of LoRA weights. If we
        # perform # pin_memory immediately after loading weights, the
        # overhead is significant.
        # 2. The weight packing above (e.g., pack_moe) may invalidate the
        # pin_memory allocation, so we execute it after packing.

        pin_memory = str(lora_device) == "cpu" and is_pin_memory_available()
        if pin_memory:
            for lora in lora_model.loras.values():
                if isinstance(lora.lora_a, list):
                    for index in range(len(lora.lora_a)):
                        if lora.lora_a[index] is None:
                            continue
                        lora.lora_a[index] = lora.lora_a[index].pin_memory()
                        lora.lora_b[index] = lora.lora_b[index].pin_memory()
                else:
                    lora.lora_a = lora.lora_a.pin_memory()
                    lora.lora_b = lora.lora_b.pin_memory()

606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
    def _stack_moe_lora_weights(
        self, lora_model: LoRAModel, module: FusedMoE3DWithLoRA, module_name: str
    ):
        module_lora = self._get_lora_layer_weights(lora_model, module_name)

        # Note (gnovack) - If MOE lora weights are not split into
        # num_experts chunks, we split them here
        if module_lora and torch.is_tensor(module_lora.lora_a):
            # Handle PEFT file format where experts.base_layer is the
            # gate_up_proj and experts is the down_proj
            gate_up_proj_lora = self._get_lora_layer_weights(
                lora_model, module_name + ".base_layer"
            )
            down_proj_lora = module_lora
            # FIXME Edge case where LoRA is not added to gate_up_proj
            # or down_proj
            assert gate_up_proj_lora is not None
            assert down_proj_lora is not None
            if self._is_3d_moe_model:
                num_experts = module.w13_lora_a_stacked[0].shape[1]

                # (num_experts,rank,input_size)
                gate_up_proj_lora.lora_a = gate_up_proj_lora.lora_a.reshape(
                    num_experts, -1, gate_up_proj_lora.lora_a.shape[-1]
                )
                down_proj_lora.lora_a = down_proj_lora.lora_a.reshape(
                    num_experts, -1, down_proj_lora.lora_a.shape[-1]
                )

                # (output_size,num_experts,rank)
                gate_up_proj_lora.lora_b = gate_up_proj_lora.lora_b.reshape(
                    gate_up_proj_lora.lora_b.shape[0], -1, num_experts
                )
                down_proj_lora.lora_b = down_proj_lora.lora_b.reshape(
                    down_proj_lora.lora_b.shape[0], -1, num_experts
                )

                # (num_experts,output_size,rank)
                gate_up_proj_lora.lora_b = gate_up_proj_lora.lora_b.permute(
                    2, 0, 1
                ).contiguous()
                down_proj_lora.lora_b = down_proj_lora.lora_b.permute(
                    2, 0, 1
                ).contiguous()

                module_lora.lora_a = [
                    gate_up_proj_lora.lora_a,
                    down_proj_lora.lora_a,
                ]
                module_lora.lora_b = [
                    gate_up_proj_lora.lora_b,
                    down_proj_lora.lora_b,
                ]
            else:
                # Some 3D MoE models haven't added the `is_3d_moe_weight`
                # attribute yet, so fallback here
                num_experts = module_lora.lora_a.shape[0] // module_lora.rank

                gate_proj_a = gate_up_proj_lora.lora_a.chunk(num_experts, dim=0)
                up_proj_a = gate_up_proj_lora.lora_a.chunk(num_experts, dim=0)

                gate_proj_b = gate_up_proj_lora.lora_b[::2, ...].chunk(
                    num_experts, dim=-1
                )
                up_proj_b = gate_up_proj_lora.lora_b[1::2, ...].chunk(
                    num_experts, dim=-1
                )

                down_proj_a = down_proj_lora.lora_a.chunk(num_experts, dim=0)
                down_proj_b = down_proj_lora.lora_b.chunk(num_experts, dim=-1)

                lora_a = []
                lora_b = []
                for i in range(num_experts):
                    lora_a.append(gate_proj_a[i])
                    lora_a.append(down_proj_a[i])
                    lora_a.append(up_proj_a[i])

                    lora_b.append(gate_proj_b[i])
                    lora_b.append(down_proj_b[i])
                    lora_b.append(up_proj_b[i])

                module_lora.lora_a = lora_a
                module_lora.lora_b = lora_b

691
    def _get_lora_layer_weights(
692
        self, lora_model: LoRAModel, module_name: str
693
    ) -> LoRALayerWeights | None:
694
        org_module_name = module_name
695
        if self.is_pooling_model and not lora_model.check_lora_name(module_name):
696
697
698
699
700
701
702
            # If it's a pool model, and the layer name is not found,
            # remove the prefix 'model.' and search again.
            module_name = module_name.replace("model.", "")
            if lora_model.check_lora_name(module_name):
                org_module_name = module_name
                logger.info_once(
                    "For the pool model, successfully loaded the LoRA weights "
703
704
                    "after removing the prefix 'model.'."
                )
705
706
        return lora_model.get_lora(org_module_name)

707
    def deactivate_adapter(self, adapter_id: int) -> bool:
708
709
710
711
712
        if adapter_id not in self._active_adapters:
            return False
        self._deactivate_adapter(adapter_id)
        self._active_adapters.pop(adapter_id, None)
        return True
713
714

    def add_adapter(self, adapter: LoRAModel) -> bool:
715
        logger.debug("Adding lora. Model id: %d, int id: %d", adapter.id, adapter.id)
716
717
718
719
720
721
        if adapter.id in self._registered_adapters:
            return False
        if len(self._registered_adapters) >= self.capacity:
            raise RuntimeError("No free adapter slots.")
        self._add_adapter(adapter)
        return True
722

723
    def set_adapter_mapping(self, mapping: LoRAMapping) -> None:
724
725
726
        if self._last_mapping != mapping:
            self._set_adapter_mapping(mapping)
            self._last_mapping = mapping
727
728

    def remove_adapter(self, adapter_id: int) -> bool:
729
730
731
732
733
        self.deactivate_adapter(adapter_id)
        if adapter_id not in self._registered_adapters:
            return False
        self._registered_adapters.pop(adapter_id, None)
        return True
734

735
736
    def list_adapters(self) -> dict[int, LoRAModel]:
        return dict(self._registered_adapters)
737

738
    def get_adapter(self, adapter_id: int) -> LoRAModel | None:
739
        return self._registered_adapters.get(adapter_id)
740
741
742


class LoRALRUCache(AdapterLRUCache[LoRAModel]):
743
    def __init__(self, capacity: int, deactivate_lora_fn: Callable[[int], bool]):
744
        super().__init__(capacity, deactivate_lora_fn)
745
746
747
748
749


class LRUCacheLoRAModelManager(LoRAModelManager):
    """A model manager that manages multiple LoRAs with LRU cache."""

750
751
752
753
754
755
756
757
    def __init__(
        self,
        model: nn.Module,
        max_num_seqs: int,
        max_num_batched_tokens: int,
        vocab_size: int,
        lora_config: LoRAConfig,
        device: torch.device,
758
        vllm_config: VllmConfig | None = None,
759
760
    ):
        super().__init__(
761
762
763
764
765
766
767
            model,
            max_num_seqs,
            max_num_batched_tokens,
            vocab_size,
            lora_config,
            device,
            vllm_config,
768
        )
769
        self._registered_adapters: LoRALRUCache = LoRALRUCache(
770
771
            self.capacity, self.deactivate_adapter
        )
772
        self._active_adapters: LoRALRUCache = LoRALRUCache(
773
774
            self.lora_slots, self._deactivate_adapter
        )
775

776
    def list_adapters(self) -> dict[int, LoRAModel]:
777
        """List all registered LoRAModels."""
778
        return dict(self._registered_adapters.cache)
779

780
    def add_adapter(self, lora: LoRAModel) -> bool:
781
        """Add a LoRAModel to the manager."""
782
        logger.debug("Adding lora. Model id: %d, int id: %d", lora.id, lora.id)
783
784
        if lora.id not in self._registered_adapters:
            self._add_adapter(lora)
785
786
787
            was_added = True
        else:
            # We always touch to update the LRU cache order
788
            self._registered_adapters.touch(lora.id)
789
790
791
            was_added = False
        return was_added

792
    def activate_adapter(
793
794
795
        self,
        lora_id: int,
    ) -> bool:
796
797
798
799
        if (
            lora_id not in self._active_adapters
            and len(self._active_adapters) >= self.lora_slots
        ):
800
801
            self._active_adapters.remove_oldest()
        result = super().activate_adapter(lora_id)
802
        # We always touch to update the LRU cache order
803
        self._active_adapters.touch(lora_id)
804
805
        return result

806
807
808
    def remove_oldest_adapter(self) -> bool:
        if len(self._registered_adapters) > 0:
            self._registered_adapters.remove_oldest()
809
810
811
            return True
        return False

812
    def pin_adapter(self, lora_id: int) -> bool:
813
814
815
816
817
818
819
        """Pin a LoRAModel in the manager cache."""
        self._pin_lora_in_cpu_cache(lora_id)
        self._pin_lora_in_gpu_cache(lora_id)
        return True

    def _pin_lora_in_cpu_cache(self, lora_id: int):
        try:
820
            self._registered_adapters.pin(lora_id)
821
        except ValueError as err:
822
823
824
            raise ValueError(
                f"Pinning failed. LoRA {lora_id} is not registered."
            ) from err
825
826

    def _pin_lora_in_gpu_cache(self, lora_id: int):
827
        if lora_id not in self._active_adapters:
828
            # move lora to gpu if not already active
829
            self.activate_adapter(lora_id)
830

831
        self._active_adapters.pin(lora_id)
832

833
834

def create_lora_manager(
835
836
837
838
839
    model: nn.Module,
    max_num_seqs: int,
    max_num_batched_tokens: int,
    vocab_size: int,
    lora_config: LoRAConfig,
840
    vllm_config: VllmConfig,
841
842
843
844
    device: torch.device,
    lora_manager_cls: type[LoRAModelManager] = LoRAModelManager,
    **kwargs,
) -> LoRAModelManager:
845
    """Create a LoRA adapter for a given model."""
846
    if not isinstance(model, SupportsLoRA):
847
848
849
850
851
852
853
        raise ValueError(f"Model {type(model)} is not supported for LoRA.")
    lora_manager = lora_manager_cls(
        model=model,
        max_num_seqs=max_num_seqs,
        max_num_batched_tokens=max_num_batched_tokens,
        vocab_size=vocab_size,
        lora_config=lora_config,
854
        vllm_config=vllm_config,
855
        device=device,
856
857
        **kwargs,
    )
858
    return lora_manager