speculative.py 25.1 KB
Newer Older
1
2
3
4
5
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

import ast
import hashlib
6
from typing import TYPE_CHECKING, Any, Literal
7

8
from pydantic import Field, SkipValidation, model_validator
9
10
11
12
13
14
from pydantic.dataclasses import dataclass
from typing_extensions import Self

from vllm.config.parallel import ParallelConfig
from vllm.config.utils import config
from vllm.logger import init_logger
15
from vllm.utils.import_utils import LazyLoader
16
17
18
19
20
21
22
23
24
25

if TYPE_CHECKING:
    from transformers import PretrainedConfig

    import vllm.model_executor.layers.quantization as me_quant
    from vllm.config import ModelConfig
else:
    PretrainedConfig = Any
    ModelConfig = Any

26
27
28
    me_quant = LazyLoader(
        "model_executor", globals(), "vllm.model_executor.layers.quantization"
    )
29
30
31

logger = init_logger(__name__)

32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
SpeculativeMethod = Literal[
    "ngram",
    "eagle",
    "eagle3",
    "medusa",
    "mlp_speculator",
    "draft_model",
    "deepseek_mtp",
    "ernie_mtp",
    "qwen3_next_mtp",
    "mimo_mtp",
    "longcat_flash_mtp",
    "mtp",
]
MTP_MODEL_TYPES = (
    "deepseek_mtp",
    "mimo_mtp",
    "glm4_moe_mtp",
    "ernie_mtp",
    "qwen3_next_mtp",
    "longcat_flash_mtp",
)
54
55
56
57
58
59


@config
@dataclass
class SpeculativeConfig:
    """Configuration for speculative decoding."""
60

61
    enforce_eager: bool | None = None
62
    """Override the default enforce_eager from model_config"""
63
    # General speculative decoding control
64
    num_speculative_tokens: int = Field(default=None, gt=0)
65
66
    """The number of speculative tokens, if provided. It will default to the
    number in the draft model config if present, otherwise, it is required."""
67
    model: str | None = None
68
69
    """The name of the draft model, eagle head, or additional weights, if
    provided."""
70
    method: SpeculativeMethod | None = None
71
72
73
74
75
76
77
    """The name of the speculative method to use. If users provide and set the
    `model` param, the speculative method type will be detected automatically
    if possible, if `model` param is not provided, the method name must be
    provided.

    If using `ngram` method, the related configuration `prompt_lookup_max` and
    `prompt_lookup_min` should be considered."""
78
    draft_tensor_parallel_size: int | None = Field(default=None, ge=1)
79
80
81
82
    """The degree of the tensor parallelism for the draft model. Can only be 1
    or the same as the target model's tensor parallel size."""

    # Draft model configuration
83
    quantization: me_quant.QuantizationMethods | None = None
84
85
86
    """Quantization method that was used to quantize the draft model weights.
    If `None`, we assume the model weights are not quantized. Note that it only
    takes effect when using the draft model-based speculative method."""
87
    max_model_len: int | None = Field(default=None, ge=1)
88
89
    """The maximum model length of the draft model. Used when testing the
    ability to skip speculation for some sequences."""
90
    revision: str | None = None
91
92
93
    """The specific model version to use for the draft model. It can be a
    branch name, a tag name, or a commit id. If unspecified, will use the
    default version."""
94
    code_revision: str | None = None
95
96
97
98
99
    """The specific revision to use for the draft model code on Hugging Face
    Hub. It can be a branch name, a tag name, or a commit id. If unspecified,
    will use the default version."""

    # Advanced control
100
    disable_by_batch_size: int | None = Field(default=None, ge=2)
101
102
    """Disable speculative decoding for new incoming requests when the number
    of enqueued requests is larger than this value, if provided."""
103
104
105
106
107
    disable_padded_drafter_batch: bool = False
    """Disable input padding for speculative decoding. If set to True,
    speculative input batches can contain sequences of different lengths,
    which may only be supported by certain attention backends. This currently
    only affects the EAGLE method of speculation."""
108
109

    # Ngram proposer configuration
110
    prompt_lookup_max: int | None = Field(default=None, ge=1)
111
112
    """Maximum size of ngram token window when using Ngram proposer, required
    when method is set to ngram."""
113
    prompt_lookup_min: int | None = Field(default=None, ge=1)
114
115
116
    """Minimum size of ngram token window when using Ngram proposer, if
    provided. Defaults to 1."""

117
    speculative_token_tree: str | None = None
118
119
120
121
122
    """Specifies the tree structure for speculative token generation.
    """
    # required configuration params passed from engine
    target_model_config: SkipValidation[ModelConfig] = None  # type: ignore
    """The configuration of the target model."""
123
    target_parallel_config: SkipValidation[ParallelConfig] = None  # type: ignore
124
125
126
127
128
    """The parallel configuration for the target model."""

    # params generated in the post-init stage
    draft_model_config: SkipValidation[ModelConfig] = None  # type: ignore
    """The configuration of the draft model initialized internal."""
129
    draft_parallel_config: SkipValidation[ParallelConfig] = None  # type: ignore
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
    """The parallel configuration for the draft model initialized internal."""

    def compute_hash(self) -> str:
        """
        WARNING: Whenever a new field is added to this config,
        ensure that it is included in the factors list if
        it affects the computation graph.

        Provide a hash that uniquely identifies all the configs
        that affect the structure of the computation
        graph from input ids/embeddings to the final hidden states,
        excluding anything before input ids/embeddings and after
        the final hidden states.
        """
        factors: list[Any] = []
        # Eagle3 affects the computation graph because it returns intermediate
        # hidden states in addition to the final hidden state.
        factors.append(self.method == "eagle3")
148
        hash_str = hashlib.md5(str(factors).encode(), usedforsecurity=False).hexdigest()
149
150
151
152
        return hash_str

    @staticmethod
    def hf_config_override(hf_config: PretrainedConfig) -> PretrainedConfig:
153
        if hf_config.model_type in ("deepseek_v3", "deepseek_v32"):
154
155
156
            hf_config.model_type = "deepseek_mtp"
        if hf_config.model_type == "deepseek_mtp":
            n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
157
158
159
            hf_config.update(
                {"n_predict": n_predict, "architectures": ["DeepSeekMTPModel"]}
            )
160
161
162
163

        if hf_config.architectures[0] == "MiMoForCausalLM":
            hf_config.model_type = "mimo_mtp"
            n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
164
165
166
167
168
169
170
            hf_config.update(
                {
                    "num_hidden_layers": 0,
                    "n_predict": n_predict,
                    "architectures": ["MiMoMTPModel"],
                }
            )
171
172
173
174

        if hf_config.architectures[0] == "Glm4MoeForCausalLM":
            hf_config.model_type = "glm4_moe_mtp"
            n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
175
176
177
178
179
180
181
            hf_config.update(
                {
                    "num_hidden_layers": 0,
                    "n_predict": n_predict,
                    "architectures": ["Glm4MoeMTPModel"],
                }
            )
182
183
184
185
186

        if hf_config.model_type == "ernie4_5_moe":
            hf_config.model_type = "ernie_mtp"
        if hf_config.model_type == "ernie_mtp":
            n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
187
188
189
            hf_config.update(
                {"n_predict": n_predict, "architectures": ["ErnieMTPModel"]}
            )
190
191
192
193
194

        if hf_config.model_type == "qwen3_next":
            hf_config.model_type = "qwen3_next_mtp"
        if hf_config.model_type == "qwen3_next_mtp":
            n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
195
196
197
            hf_config.update(
                {"n_predict": n_predict, "architectures": ["Qwen3NextMTP"]}
            )
XuruiYang's avatar
XuruiYang committed
198
199
200
        if hf_config.model_type == "longcat_flash":
            hf_config.model_type = "longcat_flash_mtp"
            n_predict = getattr(hf_config, "num_nextn_predict_layers", 1)
201
202
203
            hf_config.update(
                {"n_predict": n_predict, "architectures": ["LongCatFlashMTPModel"]}
            )
204
205
206
207
208
209
210
211
212
213
214
215

        return hf_config

    def __post_init__(self):
        # Note: "method" is a new parameter that helps to extend the
        # configuration of non-model-based proposers, and the "model" parameter
        # will be used to set the draft model, eagle head, or additional weight
        # when needed. If users do not specify "method", the speculative method
        # will be detected automatically if possible. If the speculative method
        # can not be detected, it will be considered as the "draft_model" by
        # default.

216
        if self.method in MTP_MODEL_TYPES:
217
218
219
            logger.warning(
                "method `%s` is deprecated and replaced with mtp.", self.method
            )
220
221
            self.method = "mtp"

222
        if self.model is None and self.num_speculative_tokens is not None:
223
            if self.method == "mtp":
224
225
                if self.target_model_config is None:
                    raise ValueError("target_model_config must be present for mtp")
226
                if self.target_model_config.hf_text_config.model_type == "deepseek_v32":
227
228
229
                    # FIXME(luccafong): cudgraph with v32 MTP is not supported,
                    # remove this when the issue is fixed.
                    self.enforce_eager = True
230
231
232
233
234
235
236
237
238
                # use the draft model from the same model:
                self.model = self.target_model_config.model
                # Align the quantization of draft model for cases such as
                # --quantization fp8 with a bf16 checkpoint.
                if not self.quantization:
                    self.quantization = self.target_model_config.quantization
            elif self.method in ("ngram", "[ngram]"):
                self.model = "ngram"
            else:
239
                raise ValueError(
240
241
                    "num_speculative_tokens was provided but without speculative model."
                )
242
243
244

        # Automatically configure the method for ngram when "model" is used
        # instead of "method"
245
246
247
        if self.method is None and (
            self.model is not None and self.model in ("ngram", "[ngram]")
        ):
248
249
250
251
252
253
            self.method = "ngram"

        if self.method in ("ngram", "[ngram]"):
            # Unified to "ngram" internally
            self.method = "ngram"
            # Set default values if not provided
254
            if self.prompt_lookup_min is None and self.prompt_lookup_max is None:
255
256
257
258
                # TODO(woosuk): Tune these values. They are arbitrarily chosen.
                self.prompt_lookup_min = 5
                self.prompt_lookup_max = 5
            elif self.prompt_lookup_min is None:
259
260
261
262
263
                if self.prompt_lookup_max is None:
                    raise ValueError(
                        "Either prompt_lookup_max or prompt_lookup_min must be "
                        "provided when using the ngram method."
                    )
264
265
                self.prompt_lookup_min = self.prompt_lookup_max
            elif self.prompt_lookup_max is None:
266
267
268
269
270
                if self.prompt_lookup_min is None:
                    raise ValueError(
                        "Either prompt_lookup_max or prompt_lookup_min must be "
                        "provided when using the ngram method."
                    )
271
272
273
274
275
276
                self.prompt_lookup_max = self.prompt_lookup_min

            # Validate values
            if self.prompt_lookup_min > self.prompt_lookup_max:
                raise ValueError(
                    f"prompt_lookup_min={self.prompt_lookup_min} must "
277
278
                    f"be <= prompt_lookup_max={self.prompt_lookup_max}"
                )
279
280
281
282
283
284
285
286
287
288
289
290
291
292

            # TODO: current we still need extract vocab_size from target model
            # config, in future, we may try refactor it out, and set
            # draft related config as None here.
            self.draft_model_config = self.target_model_config
            self.draft_parallel_config = self.target_parallel_config
        else:
            self.prompt_lookup_max = 0
            self.prompt_lookup_min = 0

            if self.model is not None:
                # TODO: Move this import to the top once `ModelConfig`
                # lives in `vllm.config.model`.
                from vllm.config import ModelConfig
293

294
295
296
297
298
                self.draft_model_config = ModelConfig(
                    model=self.model,
                    runner="draft",
                    tokenizer=self.target_model_config.tokenizer,
                    tokenizer_mode=self.target_model_config.tokenizer_mode,
299
300
301
                    trust_remote_code=self.target_model_config.trust_remote_code,
                    allowed_local_media_path=self.target_model_config.allowed_local_media_path,
                    allowed_media_domains=self.target_model_config.allowed_media_domains,
302
303
304
305
                    dtype=self.target_model_config.dtype,
                    seed=self.target_model_config.seed,
                    revision=self.revision,
                    code_revision=self.code_revision,
306
307
                    tokenizer_revision=self.target_model_config.tokenizer_revision,
                    spec_target_max_model_len=self.target_model_config.max_model_len,
308
309
310
311
312
313
314
                    quantization=self.quantization,
                    enforce_eager=self.target_model_config.enforce_eager,
                    max_logprobs=self.target_model_config.max_logprobs,
                    hf_overrides=SpeculativeConfig.hf_config_override,
                )

                # Automatically detect the method
315
                if self.method in ("eagle", "eagle3"):
316
317
318
319
320
321
322
323
324
325
326
                    pass
                # examples:
                # yuhuili/EAGLE-LLaMA3-Instruct-8B
                # yuhuili/EAGLE3-LLaMA3.1-Instruct-8B
                # AngelSlim/Qwen3-8B_eagle3
                elif "eagle-" in self.draft_model_config.model.lower():
                    self.method = "eagle"
                elif "eagle3" in self.draft_model_config.model.lower():
                    self.method = "eagle3"
                elif self.draft_model_config.hf_config.model_type == "medusa":
                    self.method = "medusa"
327
                elif self.draft_model_config.hf_config.model_type == "mlp_speculator":
328
                    self.method = "mlp_speculator"
329
                elif self.draft_model_config.hf_config.model_type in MTP_MODEL_TYPES:
330
                    self.method = "mtp"
331
332
                    if self.num_speculative_tokens > 1:
                        logger.warning(
333
334
335
336
337
338
339
                            "Enabling num_speculative_tokens > 1 will run"
                            "multiple times of forward on same MTP layer"
                            ",which may result in lower acceptance rate"
                        )
                elif self.draft_model_config.hf_config.model_type in (
                    "longcat_flash_mtp"
                ):
XuruiYang's avatar
XuruiYang committed
340
341
342
                    self.method = "longcat_flash_mtp"
                    if self.num_speculative_tokens > 1:
                        logger.warning(
343
344
345
346
                            "LongCat MTP models only have "
                            "one layer. Might need some code changes "
                            "to support multiple layers."
                        )
347
348
349
350
351
352
                else:
                    self.method = "draft_model"
                    raise NotImplementedError(
                        "Speculative decoding with draft model is not "
                        "supported yet. Please consider using other "
                        "speculative decoding methods such as ngram, medusa, "
353
354
                        "eagle, or mtp."
                    )
355
356
357

                # Replace hf_config for EAGLE draft_model
                if self.method in ("eagle", "eagle3"):
358
359
                    from vllm.transformers_utils.configs import SpeculatorsConfig
                    from vllm.transformers_utils.configs.eagle import EAGLEConfig
360

361
362
363
364
                    if isinstance(
                        self.draft_model_config.hf_config,
                        (EAGLEConfig, SpeculatorsConfig),
                    ):
365
366
367
368
369
                        pass
                    else:
                        eagle_config = EAGLEConfig(
                            self.draft_model_config.hf_config,
                            method=self.method,
370
371
                            model_type="eagle",
                        )
372
373
                        self.draft_model_config.hf_config = eagle_config

374
375
376
377
378
379
                if self.num_speculative_tokens is not None and hasattr(
                    self.draft_model_config.hf_config, "num_lookahead_tokens"
                ):
                    self.draft_model_config.hf_config.num_lookahead_tokens = (
                        self.num_speculative_tokens
                    )
380

381
382
383
                n_predict = getattr(
                    self.draft_model_config.hf_config, "n_predict", None
                )
384
385
386
387
                if n_predict is not None:
                    if self.num_speculative_tokens is None:
                        # Default to max value defined in draft model config.
                        self.num_speculative_tokens = n_predict
388
389
390
391
                    elif (
                        self.num_speculative_tokens > n_predict
                        and self.num_speculative_tokens % n_predict != 0
                    ):
392
393
394
                        # Ensure divisibility for MTP module reuse.
                        raise ValueError(
                            f"num_speculative_tokens:{self.num_speculative_tokens}"
395
396
                            f" must be divisible by {n_predict=}"
                        )
397
398
399

                if self.speculative_token_tree is None:
                    # Generate chain of tokens.
400
401
402
                    self.speculative_token_tree = str(
                        [(i + 1) * (0,) for i in range(self.num_speculative_tokens)]
                    )
403
404
                else:
                    # Sort the token tree breadth-first.
405
                    tree_choices = ast.literal_eval(self.speculative_token_tree)
406
                    self.speculative_token_tree = str(
407
408
                        sorted(tree_choices, key=lambda t: (len(t), t))
                    )
409

410
                self.draft_tensor_parallel_size = (
411
412
413
                    SpeculativeConfig._verify_and_get_draft_tp(
                        self.target_parallel_config,
                        self.draft_tensor_parallel_size,
414
415
                        self.draft_model_config.hf_config,
                    )
416
417
418
419
420
421
422
                )

                self.draft_model_config.max_model_len = (
                    SpeculativeConfig._maybe_override_draft_max_model_len(
                        self.max_model_len,
                        self.draft_model_config.max_model_len,
                        self.target_model_config.max_model_len,
423
424
                    )
                )
425
426
427

                self.draft_parallel_config = (
                    SpeculativeConfig.create_draft_parallel_config(
428
429
430
                        self.target_parallel_config, self.draft_tensor_parallel_size
                    )
                )
431
        return self
432
433
434

    @staticmethod
    def _maybe_override_draft_max_model_len(
435
        speculative_max_model_len: int | None,
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
        draft_max_model_len: int,
        target_max_model_len: int,
    ) -> int:
        """Determine the max sequence len for the draft model. This is usually
        the draft_max_model_len, but may be the target_max_model_len if it is
        less than the draft_max_model_len, or may be speculative_max_model_len
        if it is specified.

        This is necessary so that sequences do not exceed the capacity of the
        draft model or the target model.

        speculative_max_model_len is mainly used for testing that sequences can
        skip speculation.
        """

        if speculative_max_model_len is not None:
            if speculative_max_model_len > draft_max_model_len:
453
454
455
456
                raise ValueError(
                    f"{speculative_max_model_len=} cannot be "
                    f"larger than {draft_max_model_len=}"
                )
457
458

            if speculative_max_model_len > target_max_model_len:
459
460
461
462
                raise ValueError(
                    f"{speculative_max_model_len=} cannot be "
                    f"larger than {target_max_model_len=}"
                )
463
464
465
466
467
468
469
470
471
472

            return speculative_max_model_len

        return min(
            draft_max_model_len,
            target_max_model_len,
        )

    @staticmethod
    def _verify_and_get_draft_tp(
473
        target_parallel_config: ParallelConfig,
474
        speculative_draft_tensor_parallel_size: int | None,
475
476
        draft_hf_config: PretrainedConfig,
    ) -> int:
477
478
479
480
481
482
483
484
485
486
487
488
489
        """
        Verifies and adjusts the tensor parallel size for a draft model
        specified using speculative_draft_tensor_parallel_size.
        """
        # If speculative_draft_tensor_parallel_size is unset then set it
        # appropriately else verify that it is set correctly.
        if speculative_draft_tensor_parallel_size is None:
            if draft_hf_config.model_type == "mlp_speculator":
                speculative_draft_tensor_parallel_size = 1
                if target_parallel_config.tensor_parallel_size > 1:
                    logger.warning(
                        "%s cannot currently be run with tp>1; "
                        "setting speculative_draft_tensor_parallel_size=1",
490
491
                        draft_hf_config.model_type,
                    )
492
            else:
493
                speculative_draft_tensor_parallel_size = (
494
                    target_parallel_config.tensor_parallel_size
495
                )
496
        elif speculative_draft_tensor_parallel_size not in (
497
498
499
            1,
            target_parallel_config.tensor_parallel_size,
        ):
500
501
            raise ValueError(
                f"{speculative_draft_tensor_parallel_size=} cannot be "
502
503
                f"other value than 1 or target model tensor_parallel_size"
            )
504
505
506
507
508
509
510
511
512
513
514
515
        return speculative_draft_tensor_parallel_size

    @staticmethod
    def create_draft_parallel_config(
        target_parallel_config: ParallelConfig,
        speculative_draft_tensor_parallel_size: int,
    ) -> ParallelConfig:
        """Create a parallel config for use by the draft worker.

        This is mostly a copy of the target parallel config, except the tp_size.
        """
        draft_parallel_config = ParallelConfig(
516
            pipeline_parallel_size=target_parallel_config.pipeline_parallel_size,
517
            tensor_parallel_size=speculative_draft_tensor_parallel_size,
518
519
520
521
            distributed_executor_backend=target_parallel_config.distributed_executor_backend,
            max_parallel_loading_workers=target_parallel_config.max_parallel_loading_workers,
            disable_custom_all_reduce=target_parallel_config.disable_custom_all_reduce,
            ray_workers_use_nsight=target_parallel_config.ray_workers_use_nsight,
522
523
524
525
526
            placement_group=target_parallel_config.placement_group,
        )

        return draft_parallel_config

527
    @model_validator(mode="after")
528
529
530
531
532
    def _verify_args(self) -> Self:
        if self.num_speculative_tokens is None:
            raise ValueError(
                "num_speculative_tokens must be provided with "
                "speculative model unless the draft model config contains an "
533
534
                "n_predict parameter."
            )
535
536

        if self.num_speculative_tokens <= 0:
537
538
539
540
            raise ValueError(
                "Expected num_speculative_tokens to be greater "
                f"than zero ({self.num_speculative_tokens})."
            )
541
542
543

        if self.draft_model_config:
            self.draft_model_config.verify_with_parallel_config(
544
545
                self.draft_parallel_config
            )
546

547
548
549
550
551
552
        if self.disable_by_batch_size is not None and self.disable_by_batch_size < 2:
            raise ValueError(
                "Expect the batch size threshold of disabling "
                "speculative decoding is > 1, but got "
                f"{self.disable_by_batch_size=}"
            )
553

554
        eagle3_target_supported = ["llama", "qwen", "minicpm", "gpt_oss"]
555
556
557
558
559
560
561
562
        if (
            self.method == "eagle3"
            and self.target_model_config
            and not any(
                supported_model in self.target_model_config.hf_text_config.model_type
                for supported_model in eagle3_target_supported
            )
        ):
563
564
            raise ValueError(
                f"Eagle3 is only supported for {eagle3_target_supported} models. "  # noqa: E501
565
566
                f"Got {self.target_model_config.hf_text_config.model_type=}"
            )
567
568
569
570
571
572
573
574
575
576
577
578
579
580

        return self

    @property
    def num_lookahead_slots(self) -> int:
        """The number of additional slots the scheduler should allocate per
        step, in addition to the slots allocated for each known token.

        This is equal to the number of speculative tokens, as each speculative
        token must be scored.
        """
        return self.num_speculative_tokens

    def use_eagle(self) -> bool:
581
        return self.method in ("eagle", "eagle3", "mtp")
582
583
584
585
586
587

    def __repr__(self) -> str:
        method = self.method
        model = None if method == "ngram" else self.draft_model_config.model
        num_spec_tokens = self.num_speculative_tokens
        return f"SpeculativeConfig({method=}, {model=}, {num_spec_tokens=})"