config.py 74.9 KB
Newer Older
1
import enum
2
import json
3
from dataclasses import dataclass, field, fields
4
from typing import TYPE_CHECKING, ClassVar, List, Optional, Tuple, Type, Union
5
6

import torch
7
from transformers import PretrainedConfig
8

9
import vllm.envs as envs
Woosuk Kwon's avatar
Woosuk Kwon committed
10
from vllm.logger import init_logger
11
from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
12
from vllm.model_executor.models import ModelRegistry
13
from vllm.tracing import is_otel_installed
14
from vllm.transformers_utils.config import get_config, get_hf_text_config
15
from vllm.utils import (STR_NOT_IMPL_ENC_DEC_CUDAGRAPH, GiB_bytes,
16
                        cuda_device_count_stateless, get_cpu_memory, is_cpu,
17
                        is_hip, is_neuron, is_openvino, is_tpu, is_xpu,
18
                        print_warning_once)
19

20
21
22
if TYPE_CHECKING:
    from ray.util.placement_group import PlacementGroup

23
    from vllm.executor.executor_base import ExecutorBase
24
    from vllm.model_executor.model_loader.loader import BaseModelLoader
25
26
    from vllm.transformers_utils.tokenizer_group.base_tokenizer_group import (
        BaseTokenizerGroup)
27

28
29
logger = init_logger(__name__)

30
_EMBEDDING_MODEL_MAX_NUM_BATCHED_TOKENS = 32768
31

32
33
34
_PP_SUPPORTED_MODELS = [
    "AquilaModel",
    "AquilaForCausalLM",
35
    "DeepseekV2ForCausalLM",
36
37
38
39
40
41
    "InternLMForCausalLM",
    "LlamaForCausalLM",
    "LLaMAForCausalLM",
    "MistralForCausalLM",
    "Phi3ForCausalLM",
    "GPT2LMHeadModel",
42
    "MixtralForCausalLM",
43
    "NemotronForCausalLM",
44
45
    "Qwen2ForCausalLM",
    "Qwen2MoeForCausalLM",
46
    "QWenLMHeadModel",
47
48
]

49
50

class ModelConfig:
51
52
53
54
    """Configuration for the model.

    Args:
        model: Name or path of the huggingface model to use.
55
56
            It is also used as the content for `model_name` tag in metrics 
            output when `served_model_name` is not specified. 
57
        tokenizer: Name or path of the huggingface tokenizer to use.
58
59
        tokenizer_mode: Tokenizer mode. "auto" will use the fast tokenizer if
            available, and "slow" will always use the slow tokenizer.
60
61
        trust_remote_code: Trust remote code (e.g., from HuggingFace) when
            downloading the model and tokenizer.
62
63
64
65
        dtype: Data type for model weights and activations. The "auto" option
            will use FP16 precision for FP32 and FP16 models, and BF16 precision
            for BF16 models.
        seed: Random seed for reproducibility.
Jasmond L's avatar
Jasmond L committed
66
67
68
        revision: The specific model version to use. It can be a branch name,
            a tag name, or a commit id. If unspecified, will use the default
            version.
69
        code_revision: The specific revision to use for the model code on
70
            Hugging Face Hub. It can be a branch name, a tag name, or a
71
            commit id. If unspecified, will use the default version.
72
73
74
        rope_scaling: Dictionary containing the scaling configuration for the
            RoPE embeddings. When using this flag, don't update
            `max_position_embeddings` to the expected new maximum.
75
76
77
        tokenizer_revision: The specific tokenizer version to use. It can be a
            branch name, a tag name, or a commit id. If unspecified, will use
            the default version.
78
79
        max_model_len: Maximum length of a sequence (including prompt and
            output). If None, will be derived from the model.
80
81
        quantization: Quantization method that was used to quantize the model
            weights. If None, we assume the model weights are not quantized.
82
83
        quantization_param_path: Path to JSON file containing scaling factors.
            Used to load KV cache scaling factors into the model when KV cache
84
85
            type is FP8_E4M3 on ROCm (AMD GPU). In the future these will also
            be used to load activation and weight scaling factors when the
86
            model dtype is FP8_E4M3 on ROCm.
87
88
89
        enforce_eager: Whether to enforce eager execution. If True, we will
            disable CUDA graph and always execute the model in eager mode.
            If False, we will use CUDA graph and eager execution in hybrid.
90
91
92
            If None, the user did not specify, so default to False -
            except for encoder/decoder models, which currently require
            eager mode.
93
94
        max_context_len_to_capture: Maximum context len covered by CUDA graphs.
            When a sequence has context length larger than this, we fall back
95
96
97
98
            to eager mode (DEPRECATED. Use max_seq_len_to_capture instead).
        max_seq_len_to_capture: Maximum sequence len covered by CUDA graphs.
            When a sequence has context length larger than this, we fall back
            to eager mode
99
100
101
102
        disable_sliding_window: Whether to disable sliding window. If True,
            we will disable the sliding window functionality of the model.
            If the model does not support sliding window, this argument is
            ignored.
103
104
        skip_tokenizer_init: If true, skip initialization of tokenizer and
            detokenizer.
105
106
107
108
        served_model_name: The model name used in metrics tag `model_name`,
            matches the model name exposed via the APIs. If multiple model 
            names provided, the first name will be used. If not specified, 
            the model name will be the same as `model`.
109
    """
110
111
112
113

    def __init__(
        self,
        model: str,
114
115
        tokenizer: str,
        tokenizer_mode: str,
116
        trust_remote_code: bool,
117
        dtype: Union[str, torch.dtype],
118
        seed: int,
119
        revision: Optional[str] = None,
120
        code_revision: Optional[str] = None,
121
        rope_scaling: Optional[dict] = None,
122
        rope_theta: Optional[float] = None,
123
        tokenizer_revision: Optional[str] = None,
124
        max_model_len: Optional[int] = None,
125
        quantization: Optional[str] = None,
126
        quantization_param_path: Optional[str] = None,
127
        enforce_eager: Optional[bool] = None,
128
        max_context_len_to_capture: Optional[int] = None,
129
        max_seq_len_to_capture: Optional[int] = None,
130
        max_logprobs: int = 20,
131
        disable_sliding_window: bool = False,
132
        skip_tokenizer_init: bool = False,
133
        served_model_name: Optional[Union[str, List[str]]] = None,
134
        multimodal_config: Optional["MultiModalConfig"] = None,
135
136
    ) -> None:
        self.model = model
137
        self.tokenizer = tokenizer
138
        self.tokenizer_mode = tokenizer_mode
139
        self.trust_remote_code = trust_remote_code
140
        self.seed = seed
Jasmond L's avatar
Jasmond L committed
141
        self.revision = revision
142
        self.code_revision = code_revision
143
        self.rope_scaling = rope_scaling
144
        self.rope_theta = rope_theta
145
146
147
148
149
        # The tokenizer version is consistent with the model version by default.
        if tokenizer_revision is None:
            self.tokenizer_revision = revision
        else:
            self.tokenizer_revision = tokenizer_revision
150
        self.quantization = quantization
151
        self.quantization_param_path = quantization_param_path
152
        self.enforce_eager = enforce_eager
153
        if max_context_len_to_capture is not None:
154
155
            raise ValueError("`max_context_len_to_capture` is deprecated. "
                             "Use `max_seq_len_to_capture` instead.")
156
        self.max_seq_len_to_capture = max_seq_len_to_capture
157
        self.max_logprobs = max_logprobs
158
        self.disable_sliding_window = disable_sliding_window
159
        self.skip_tokenizer_init = skip_tokenizer_init
160

161
        self.hf_config = get_config(self.model, trust_remote_code, revision,
162
                                    code_revision, rope_scaling, rope_theta)
163
164
        self.hf_text_config = get_hf_text_config(self.hf_config)
        self.dtype = _get_and_verify_dtype(self.hf_text_config, dtype)
Woosuk Kwon's avatar
Woosuk Kwon committed
165

166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
        # Choose a default enforce_eager value if the user did not specify
        # a value (enforce_eager is None)
        if getattr(self.hf_config, 'is_encoder_decoder', False):
            if self.enforce_eager is None:
                # *Only for encoder/decoder models* and
                # *only if enforce_eager is unset*, override
                # to enforce_eager=True
                #
                # Add a logger message since it is *somewhat* non-intuitive that
                # enforce_eager is True when the user has not specified its
                # value.
                logger.info("Forcing enforce_eager == True because "
                            "enforce_eager setting was unspecified and "
                            "CUDAGraph is not supported with encoder/ "
                            "decoder models.")
                self.enforce_eager = True

            if not self.enforce_eager:
                # Eager mode explicitly disabled by user for an encoder/
                # decoder model; however CUDAGRAPH + encoder/decoder is
                # not currently supported
                raise ValueError(STR_NOT_IMPL_ENC_DEC_CUDAGRAPH)
        elif self.enforce_eager is None:
            # *Only for decoder-only models*, enforce_eager
            # defaults to False if unset. This is intuitive
            # so no logging message needed.
            self.enforce_eager = False

Woosuk Kwon's avatar
Woosuk Kwon committed
194
195
196
197
198
199
200
201
202
203
        if (not self.disable_sliding_window
                and self.hf_text_config.model_type == "gemma2"
                and self.hf_text_config.sliding_window is not None):
            print_warning_once(
                "Gemma 2 uses sliding window attention for every odd layer, "
                "which is currently not supported by vLLM. Disabling sliding "
                "window and capping the max length to the sliding window size "
                f"({self.hf_text_config.sliding_window}).")
            self.disable_sliding_window = True

204
205
206
207
208
        self.max_model_len = _get_and_verify_max_len(
            hf_config=self.hf_text_config,
            max_model_len=max_model_len,
            disable_sliding_window=self.disable_sliding_window,
            sliding_window_len=self.get_hf_config_sliding_window())
209
210
        self.served_model_name = get_served_model_name(model,
                                                       served_model_name)
211
212
        self.multimodal_config = multimodal_config

213
214
        if not self.skip_tokenizer_init:
            self._verify_tokenizer_mode()
215
        self._verify_embedding_mode()
216
        self._verify_quantization()
217
        self._verify_cuda_graph()
218
219
220
221
222
223
224
225

    def _verify_tokenizer_mode(self) -> None:
        tokenizer_mode = self.tokenizer_mode.lower()
        if tokenizer_mode not in ["auto", "slow"]:
            raise ValueError(
                f"Unknown tokenizer mode: {self.tokenizer_mode}. Must be "
                "either 'auto' or 'slow'.")
        self.tokenizer_mode = tokenizer_mode
226

227
228
229
230
231
    def _verify_embedding_mode(self) -> None:
        architectures = getattr(self.hf_config, "architectures", [])
        self.embedding_mode = any(
            ModelRegistry.is_embedding_model(arch) for arch in architectures)

232
233
234
    def _parse_quant_hf_config(self):
        quant_cfg = getattr(self.hf_config, "quantization_config", None)
        if quant_cfg is None:
235
            # compressed-tensors uses a "compression_config" key
236
            quant_cfg = getattr(self.hf_config, "compression_config", None)
237
238
        return quant_cfg

239
    def _verify_quantization(self) -> None:
240
241
        supported_quantization = [*QUANTIZATION_METHODS]
        rocm_supported_quantization = ["gptq", "squeezellm"]
242
243
244
245
        optimized_quantization_methods = [
            "fp8", "marlin", "gptq_marlin_24", "gptq_marlin", "awq_marlin",
            "fbgemm_fp8", "compressed_tensors", "compressed-tensors"
        ]
246
        tpu_supported_quantization = ["tpu_int8"]
247
248
249
250
        if self.quantization is not None:
            self.quantization = self.quantization.lower()

        # Parse quantization method from the HF model config, if available.
251
252
        quant_cfg = self._parse_quant_hf_config()

253
254
        if quant_cfg is not None:
            quant_method = quant_cfg.get("quant_method", "").lower()
255
256

            # Detect which checkpoint is it
257
            for _, method in QUANTIZATION_METHODS.items():
258
259
260
261
262
263
                quantization_override = method.override_quantization_method(
                    quant_cfg, self.quantization)
                if quantization_override:
                    quant_method = quantization_override
                    self.quantization = quantization_override
                    break
264

265
            # Verify quantization configurations.
266
            if self.quantization is None:
267
268
                self.quantization = quant_method
            elif self.quantization != quant_method:
269
270
                raise ValueError(
                    "Quantization method specified in the model config "
271
                    f"({quant_method}) does not match the quantization "
272
273
274
275
276
277
278
279
                    f"method specified in the `quantization` argument "
                    f"({self.quantization}).")

        if self.quantization is not None:
            if self.quantization not in supported_quantization:
                raise ValueError(
                    f"Unknown quantization method: {self.quantization}. Must "
                    f"be one of {supported_quantization}.")
280
            if is_hip(
281
            ) and self.quantization not in rocm_supported_quantization:
282
                raise ValueError(
283
284
                    f"{self.quantization} quantization is currently not "
                    f"supported in ROCm.")
285
286
287
288
289
            if is_tpu(
            ) and self.quantization not in tpu_supported_quantization:
                raise ValueError(
                    f"{self.quantization} quantization is currently not "
                    f"supported in TPU Backend.")
290
            if self.quantization not in optimized_quantization_methods:
291
                logger.warning(
292
                    "%s quantization is not fully "
293
                    "optimized yet. The speed can be slower than "
294
                    "non-quantized models.", self.quantization)
295

296
    def _verify_cuda_graph(self) -> None:
297
298
299
300
        if self.max_seq_len_to_capture is None:
            self.max_seq_len_to_capture = self.max_model_len
        self.max_seq_len_to_capture = min(self.max_seq_len_to_capture,
                                          self.max_model_len)
301

302
303
304
305
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
306
307
        total_num_attention_heads = getattr(self.hf_text_config,
                                            "num_attention_heads", 0)
308
309
310
311
312
313
314
315
        tensor_parallel_size = parallel_config.tensor_parallel_size
        if total_num_attention_heads % tensor_parallel_size != 0:
            raise ValueError(
                f"Total number of attention heads ({total_num_attention_heads})"
                " must be divisible by tensor parallel size "
                f"({tensor_parallel_size}).")

        pipeline_parallel_size = parallel_config.pipeline_parallel_size
316
317
318
319
320
321
322
        architectures = getattr(self.hf_config, "architectures", [])
        if not all(arch in _PP_SUPPORTED_MODELS
                   for arch in architectures) and pipeline_parallel_size > 1:
            raise NotImplementedError(
                "Pipeline parallelism is only supported for the following "
                f" architectures: {_PP_SUPPORTED_MODELS}.")

323
324
325
326
327
328
        if self.quantization == "bitsandbytes" and (
                parallel_config.tensor_parallel_size > 1
                or parallel_config.pipeline_parallel_size > 1):
            raise ValueError(
                "BitAndBytes quantization with TP or PP is not supported yet.")

329
        if self.quantization == "bitsandbytes" and self.enforce_eager is False:
330
331
332
            logger.warning("CUDA graph is not supported on BitAndBytes yet, "
                           "fallback to the eager mode.")
            self.enforce_eager = True
333

334
    def get_hf_config_sliding_window(self) -> Optional[int]:
Woosuk Kwon's avatar
Woosuk Kwon committed
335
        """Get the sliding window size, or None if disabled."""
336
337
338
339

        # Some models, like Qwen2 and Qwen1.5, use `use_sliding_window` in
        # addition to sliding window size. We check if that field is present
        # and if it's False, return None.
340
341
        if (hasattr(self.hf_text_config, "use_sliding_window")
                and not self.hf_text_config.use_sliding_window):
342
            return None
343
        return getattr(self.hf_text_config, "sliding_window", None)
344

345
346
347
348
349
350
351
352
353
    def get_sliding_window(self) -> Optional[int]:
        """Get the sliding window size, or None if disabled.
        """
        # If user disables sliding window, return None.
        if self.disable_sliding_window:
            return None
        # Otherwise get the value from the hf config.
        return self.get_hf_config_sliding_window()

354
    def get_vocab_size(self) -> int:
355
        return self.hf_text_config.vocab_size
356

357
    def get_hidden_size(self) -> int:
358
        return self.hf_text_config.hidden_size
359
360

    def get_head_size(self) -> int:
wangding zeng's avatar
wangding zeng committed
361
362
363
364
365
366
        # TODO remove hard code
        if hasattr(self.hf_text_config, "model_type"
                   ) and self.hf_text_config.model_type == 'deepseek_v2':
            # FlashAttention supports only head_size 32, 64, 128, 256,
            # we need to pad head_size 192 to 256
            return 256
367
368
        if hasattr(self.hf_text_config, "head_dim"):
            return self.hf_text_config.head_dim
369
        # FIXME(woosuk): This may not be true for all models.
370
371
        return (self.hf_text_config.hidden_size //
                self.hf_text_config.num_attention_heads)
372

373
374
    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
Zhuohan Li's avatar
Zhuohan Li committed
375
        # For GPTBigCode & Falcon:
376
        # NOTE: for falcon, when new_decoder_architecture is True, the
Zhuohan Li's avatar
Zhuohan Li committed
377
378
        # multi_query flag is ignored and we use n_head_kv for the number of
        # KV heads.
379
        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
380
        new_decoder_arch_falcon = (
381
            self.hf_config.model_type in falcon_model_types
382
            and getattr(self.hf_config, "new_decoder_architecture", False))
383
        if not new_decoder_arch_falcon and getattr(self.hf_text_config,
384
                                                   "multi_query", False):
Zhuohan Li's avatar
Zhuohan Li committed
385
            # Multi-query attention, only one KV head.
Woosuk Kwon's avatar
Woosuk Kwon committed
386
            # Currently, tensor parallelism is not supported in this case.
Zhuohan Li's avatar
Zhuohan Li committed
387
            return 1
388

389
        # For DBRX and MPT
390
391
392
393
394
        if self.hf_config.model_type == "mpt":
            if "kv_n_heads" in self.hf_config.attn_config:
                return self.hf_config.attn_config["kv_n_heads"]
            return self.hf_config.num_attention_heads
        if self.hf_config.model_type == "dbrx":
395
396
397
            return getattr(self.hf_config.attn_config, "kv_n_heads",
                           self.hf_config.num_attention_heads)

398
399
400
401
402
403
404
405
406
407
        attributes = [
            # For Falcon:
            "n_head_kv",
            "num_kv_heads",
            # For LLaMA-2:
            "num_key_value_heads",
            # For ChatGLM:
            "multi_query_group_num",
        ]
        for attr in attributes:
408
            num_kv_heads = getattr(self.hf_text_config, attr, None)
409
410
411
412
413
            if num_kv_heads is not None:
                return num_kv_heads

        # For non-grouped-query attention models, the number of KV heads is
        # equal to the number of attention heads.
414
        return self.hf_text_config.num_attention_heads
415
416
417
418
419
420
421
422
423
424

    def get_num_kv_heads(self, parallel_config: "ParallelConfig") -> int:
        """Returns the number of KV heads per GPU."""
        total_num_kv_heads = self.get_total_num_kv_heads()
        # If tensor parallelism is used, we divide the number of KV heads by
        # the tensor parallel size. We will replicate the KV heads in the
        # case where the number of KV heads is smaller than the tensor
        # parallel size so each GPU has at least one KV head.
        return max(1,
                   total_num_kv_heads // parallel_config.tensor_parallel_size)
425

426
427
    def get_num_attention_heads(self,
                                parallel_config: "ParallelConfig") -> int:
428
429
        num_heads = getattr(self.hf_text_config, "num_attention_heads", 0)
        return num_heads // parallel_config.tensor_parallel_size
430

431
    def get_num_layers(self, parallel_config: "ParallelConfig") -> int:
432
        from vllm.distributed.utils import get_pp_indices
Mor Zusman's avatar
Mor Zusman committed
433
434
        total_num_hidden_layers = getattr(self.hf_text_config,
                                          "num_hidden_layers", 0)
435
436
437
438
        pp_rank = parallel_config.rank // parallel_config.tensor_parallel_size
        pp_size = parallel_config.pipeline_parallel_size
        start, end = get_pp_indices(total_num_hidden_layers, pp_rank, pp_size)
        return end - start
439

Mor Zusman's avatar
Mor Zusman committed
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
    def contains_seqlen_agnostic_layers(
            self, parallel_config: "ParallelConfig") -> bool:
        """True for Mamba/SSM models (Jamba)"""
        return self._get_num_seqlen_agnostic_layers(parallel_config) > 0

    def get_layers_block_type(self,
                              parallel_config: "ParallelConfig") -> List[str]:
        num_layers = self.get_num_layers(parallel_config)
        # Transformers supports layers_block_type @property
        return getattr(self.hf_config, "layers_block_type",
                       ["attention"] * num_layers)

    def get_num_attention_layers(self,
                                 parallel_config: "ParallelConfig") -> int:
        return len([
            t for t in self.get_layers_block_type(parallel_config)
            if t == "attention"
        ])

    def _get_num_seqlen_agnostic_layers(
            self, parallel_config: "ParallelConfig") -> int:
        return len([
            t for t in self.get_layers_block_type(parallel_config)
            if t != "attention"
        ])

466
467
468
469
470
471
472
473
474
475
    @property
    def is_encoder_decoder_model(self) -> bool:
        """Extract the HF encoder/decoder model flag."""
        return getattr(self.hf_config, "is_encoder_decoder", False)

    @property
    def is_embedding_model(self) -> bool:
        """Extract the embedding model flag."""
        return self.embedding_mode

476
477

class CacheConfig:
478
479
480
481
482
    """Configuration for the KV cache.

    Args:
        block_size: Size of a cache block in number of tokens.
        gpu_memory_utilization: Fraction of GPU memory to use for the
Woosuk Kwon's avatar
Woosuk Kwon committed
483
            vLLM execution.
484
        swap_space: Size of the CPU swap space per GPU (in GiB).
485
        cache_dtype: Data type for kv cache storage.
486
        num_gpu_blocks_override: Number of GPU blocks to use. This overrides the
487
            profiled num_gpu_blocks if specified. Does nothing if None.
488
    """
489

490
491
492
493
    def __init__(
        self,
        block_size: int,
        gpu_memory_utilization: float,
494
        swap_space: float,
495
        cache_dtype: str,
496
        num_gpu_blocks_override: Optional[int] = None,
497
        sliding_window: Optional[int] = None,
498
        enable_prefix_caching: bool = False,
499
        cpu_offload_gb: float = 0,
500
501
502
    ) -> None:
        self.block_size = block_size
        self.gpu_memory_utilization = gpu_memory_utilization
503
        self.swap_space_bytes = swap_space * GiB_bytes
504
        self.num_gpu_blocks_override = num_gpu_blocks_override
505
        self.cache_dtype = cache_dtype
506
        self.sliding_window = sliding_window
507
        self.enable_prefix_caching = enable_prefix_caching
508
        self.cpu_offload_gb = cpu_offload_gb
509
        self._verify_args()
510
        self._verify_cache_dtype()
511
        self._verify_prefix_caching()
512
513
514
515
516

        # Will be set after profiling.
        self.num_gpu_blocks = None
        self.num_cpu_blocks = None

517
    def metrics_info(self):
518
519
        # convert cache_config to dict(key: str, value: str) for prometheus
        # metrics info
520
521
        return {key: str(value) for key, value in self.__dict__.items()}

522
523
524
525
526
527
    def _verify_args(self) -> None:
        if self.gpu_memory_utilization > 1.0:
            raise ValueError(
                "GPU memory utilization must be less than 1.0. Got "
                f"{self.gpu_memory_utilization}.")

528
529
530
    def _verify_cache_dtype(self) -> None:
        if self.cache_dtype == "auto":
            pass
531
        elif self.cache_dtype in ("fp8", "fp8_e4m3", "fp8_e5m2"):
532
            logger.info(
533
534
                "Using fp8 data type to store kv cache. It reduces the GPU "
                "memory footprint and boosts the performance. "
535
536
                "Meanwhile, it may cause accuracy drop without a proper "
                "scaling factor")
537
538
539
        else:
            raise ValueError(f"Unknown kv cache dtype: {self.cache_dtype}")

540
541
542
543
544
545
546
547
548
    def _verify_prefix_caching(self) -> None:
        if not self.enable_prefix_caching:
            return

        if self.sliding_window is not None:
            raise NotImplementedError(
                "Prefix caching is not supported with sliding window. "
                "Run with --disable-sliding-window to use prefix caching.")

549
550
551
552
553
554
555
556
557
558
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
        total_cpu_memory = get_cpu_memory()
        # FIXME(woosuk): Here, it is assumed that the GPUs in a tensor parallel
        # group are in the same node. However, the GPUs may span multiple nodes.
        num_gpus_per_node = parallel_config.tensor_parallel_size
        cpu_memory_usage = self.swap_space_bytes * num_gpus_per_node

559
560
561
        msg = (f"{cpu_memory_usage / GiB_bytes:.2f} GiB out of the "
               f"{total_cpu_memory / GiB_bytes:.2f} GiB total CPU memory "
               "is allocated for the swap space.")
562
563
564
        if cpu_memory_usage > 0.7 * total_cpu_memory:
            raise ValueError("Too large swap space. " + msg)
        elif cpu_memory_usage > 0.4 * total_cpu_memory:
565
            logger.warning("Possibly too large swap space. %s", msg)
566

567

568
569
570
@dataclass
class TokenizerPoolConfig:
    """Configuration for the tokenizer pool.
571

572
573
574
575
576
577
578
579
    Args:
        pool_size: Number of tokenizer workers in the pool.
        pool_type: Type of the pool.
        extra_config: Additional config for the pool.
            The way the config will be used depends on the
            pool type.
    """
    pool_size: int
580
    pool_type: Union[str, Type["BaseTokenizerGroup"]]
581
582
583
    extra_config: dict

    def __post_init__(self):
584
585
        if self.pool_type not in ("ray", ) and not isinstance(
                self.pool_type, type):
586
587
588
589
590
591
592
593
594
595
            raise ValueError(f"Unknown pool type: {self.pool_type}")
        if not isinstance(self.extra_config, dict):
            raise ValueError("extra_config must be a dictionary.")

    @classmethod
    def create_config(
        cls, tokenizer_pool_size: int, tokenizer_pool_type: str,
        tokenizer_pool_extra_config: Optional[Union[str, dict]]
    ) -> Optional["TokenizerPoolConfig"]:
        """Create a TokenizerPoolConfig from the given parameters.
596

597
        If tokenizer_pool_size is 0, return None.
598

599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
        Args:
            tokenizer_pool_size: Number of tokenizer workers in the pool.
            tokenizer_pool_type: Type of the pool.
            tokenizer_pool_extra_config: Additional config for the pool.
                The way the config will be used depends on the
                pool type. This can be a JSON string (will be parsed).
        """
        if tokenizer_pool_size:
            if isinstance(tokenizer_pool_extra_config, str):
                tokenizer_pool_extra_config_parsed = json.loads(
                    tokenizer_pool_extra_config)
            else:
                tokenizer_pool_extra_config_parsed = (
                    tokenizer_pool_extra_config or {})
            tokenizer_pool_config = cls(tokenizer_pool_size,
                                        tokenizer_pool_type,
                                        tokenizer_pool_extra_config_parsed)
        else:
            tokenizer_pool_config = None
        return tokenizer_pool_config


621
622
623
624
625
626
627
class LoadFormat(str, enum.Enum):
    AUTO = "auto"
    PT = "pt"
    SAFETENSORS = "safetensors"
    NPCACHE = "npcache"
    DUMMY = "dummy"
    TENSORIZER = "tensorizer"
628
    SHARDED_STATE = "sharded_state"
629
    GGUF = "gguf"
630
    BITSANDBYTES = "bitsandbytes"
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649


@dataclass
class LoadConfig:
    """
        download_dir: Directory to download and load the weights, default to the
            default cache directory of huggingface.
        load_format: The format of the model weights to load:
            "auto" will try to load the weights in the safetensors format and
                fall back to the pytorch bin format if safetensors format is
                not available.
            "pt" will load the weights in the pytorch bin format.
            "safetensors" will load the weights in the safetensors format.
            "npcache" will load the weights in pytorch format and store
                a numpy cache to speed up the loading.
            "dummy" will initialize the weights with random values, which is
                mainly for profiling.
            "tensorizer" will use CoreWeave's tensorizer library for
                fast weight loading.
650
            "bitsandbytes" will load nf4 type weights.
651
652
653
        ignore_patterns: The list of patterns to ignore when loading the model.
            Default to "original/**/*" to avoid repeated loading of llama's 
            checkpoints.
654
            
655
656
657
658
659
660
    """

    load_format: Union[str, LoadFormat, "BaseModelLoader"] = LoadFormat.AUTO
    download_dir: Optional[str] = None
    model_loader_extra_config: Optional[Union[str, dict]] = field(
        default_factory=dict)
661
    ignore_patterns: Optional[Union[List[str], str]] = None
662
663
664
665
666
667
668
669

    def __post_init__(self):
        model_loader_extra_config = self.model_loader_extra_config or {}
        if isinstance(model_loader_extra_config, str):
            self.model_loader_extra_config = json.loads(
                model_loader_extra_config)
        self._verify_load_format()

670
671
672
673
674
675
676
        if self.ignore_patterns is not None and len(self.ignore_patterns) > 0:
            logger.info(
                "Ignoring the following patterns when downloading weights: %s",
                self.ignore_patterns)
        else:
            self.ignore_patterns = ["original/**/*"]

677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
    def _verify_load_format(self) -> None:
        if not isinstance(self.load_format, str):
            return

        load_format = self.load_format.lower()
        self.load_format = LoadFormat(load_format)

        rocm_not_supported_load_format: List[str] = []
        if is_hip() and load_format in rocm_not_supported_load_format:
            rocm_supported_load_format = [
                f for f in LoadFormat.__members__
                if (f not in rocm_not_supported_load_format)
            ]
            raise ValueError(
                f"load format '{load_format}' is not supported in ROCm. "
                f"Supported load formats are "
                f"{rocm_supported_load_format}")


696
class ParallelConfig:
697
698
699
700
701
    """Configuration for the distributed execution.

    Args:
        pipeline_parallel_size: Number of pipeline parallel groups.
        tensor_parallel_size: Number of tensor parallel groups.
702
        worker_use_ray: Deprecated, use distributed_executor_backend instead.
zspo's avatar
zspo committed
703
704
705
        max_parallel_loading_workers: Maximum number of multiple batches
            when load model sequentially. To avoid RAM OOM when using tensor
            parallel and large models.
706
707
        disable_custom_all_reduce: Disable the custom all-reduce kernel and
            fall back to NCCL.
708
709
        tokenizer_pool_config: Config for the tokenizer pool.
            If None, will use synchronous tokenization.
710
711
        ray_workers_use_nsight: Whether to profile Ray workers with nsight, see
            https://docs.ray.io/en/latest/ray-observability/user-guides/profiling.html#profiling-nsight-profiler.
712
        placement_group: ray distributed model workers placement group.
713
714
715
716
        distributed_executor_backend: Backend to use for distributed model
            workers, either "ray" or "mp" (multiprocessing). If either
            pipeline_parallel_size or tensor_parallel_size is greater than 1,
            will default to "ray" if Ray is installed or "mp" otherwise.
717
    """
718

719
720
721
722
    def __init__(
        self,
        pipeline_parallel_size: int,
        tensor_parallel_size: int,
723
        worker_use_ray: Optional[bool] = None,
724
        max_parallel_loading_workers: Optional[int] = None,
725
        disable_custom_all_reduce: bool = False,
726
        tokenizer_pool_config: Optional[TokenizerPoolConfig] = None,
727
        ray_workers_use_nsight: bool = False,
728
        placement_group: Optional["PlacementGroup"] = None,
729
730
        distributed_executor_backend: Optional[Union[
            str, Type["ExecutorBase"]]] = None,
731
732
    ) -> None:
        self.pipeline_parallel_size = pipeline_parallel_size
733
        self.tensor_parallel_size = tensor_parallel_size
734
        self.distributed_executor_backend = distributed_executor_backend
735
        self.max_parallel_loading_workers = max_parallel_loading_workers
736
        self.disable_custom_all_reduce = disable_custom_all_reduce
737
        self.tokenizer_pool_config = tokenizer_pool_config
738
        self.ray_workers_use_nsight = ray_workers_use_nsight
739
        self.placement_group = placement_group
740

741
        self.world_size = pipeline_parallel_size * self.tensor_parallel_size
742
743
744
        if worker_use_ray:
            if self.distributed_executor_backend is None:
                self.distributed_executor_backend = "ray"
745
            elif not self.use_ray:
746
747
748
749
750
                raise ValueError(f"worker-use-ray can't be used with "
                                 f"distributed executor backend "
                                 f"'{self.distributed_executor_backend}'.")

        if self.distributed_executor_backend is None and self.world_size > 1:
751
752
753
            # We use multiprocessing by default if world_size fits on the
            # current node and we aren't in a ray placement group.

754
            from vllm.executor import ray_utils
755
            backend = "mp"
756
            ray_found = ray_utils.ray_is_available()
757
            if cuda_device_count_stateless() < self.world_size:
758
759
                if not ray_found:
                    raise ValueError("Unable to load Ray which is "
760
761
762
                                     "required for multi-node inference, "
                                     "please install Ray with `pip install "
                                     "ray`.") from ray_utils.ray_import_err
763
764
                backend = "ray"
            elif ray_found:
765
                if self.placement_group:
766
                    backend = "ray"
767
768
769
770
771
772
                else:
                    from ray import is_initialized as ray_is_initialized
                    if ray_is_initialized():
                        from ray.util import get_current_placement_group
                        if get_current_placement_group():
                            backend = "ray"
773
774
775
            self.distributed_executor_backend = backend
            logger.info("Defaulting to use %s for distributed inference",
                        backend)
776

777
        self._verify_args()
778
        self.rank: int = 0
779

780
781
782
783
784
785
    @property
    def use_ray(self) -> bool:
        return self.distributed_executor_backend == "ray" or (
            isinstance(self.distributed_executor_backend, type)
            and self.distributed_executor_backend.uses_ray)

786
    def _verify_args(self) -> None:
787
788
789
790
791
792
793
        # Lazy import to avoid circular import
        from vllm.executor.executor_base import ExecutorBase

        if self.distributed_executor_backend not in (
                "ray", "mp", None) and not (isinstance(
                    self.distributed_executor_backend, type) and issubclass(
                        self.distributed_executor_backend, ExecutorBase)):
794
            raise ValueError(
795
796
797
798
                "Unrecognized distributed executor backend "
                f"{self.distributed_executor_backend}. Supported "
                "values are 'ray', 'mp' or custom ExecutorBase subclass.")
        if self.use_ray:
799
800
            from vllm.executor import ray_utils
            ray_utils.assert_ray_available()
801
802
803
804
805
        if is_hip():
            self.disable_custom_all_reduce = True
            logger.info(
                "Disabled the custom all-reduce kernel because it is not "
                "supported on AMD GPUs.")
806
        if self.ray_workers_use_nsight and not self.use_ray:
807
808
            raise ValueError("Unable to use nsight profiling unless workers "
                             "run with Ray.")
809

810
811

class SchedulerConfig:
812
813
814
815
816
817
818
    """Scheduler configuration.

    Args:
        max_num_batched_tokens: Maximum number of tokens to be processed in
            a single iteration.
        max_num_seqs: Maximum number of sequences to be processed in a single
            iteration.
Chaofan Lin's avatar
Chaofan Lin committed
819
        max_model_len: Maximum length of a sequence (including prompt
Lily Liu's avatar
Lily Liu committed
820
            and generated text).
821
822
823
824
825
        use_v2_block_manager: Whether to use the BlockSpaceManagerV2 or not.
        num_lookahead_slots: The number of slots to allocate per sequence per
            step, beyond the known token ids. This is used in speculative
            decoding to store KV activations of tokens which may or may not be
            accepted.
826
827
        delay_factor: Apply a delay (of delay factor multiplied by previous
            prompt latency) before scheduling next prompt.
828
829
        enable_chunked_prefill: If True, prefill requests can be chunked based
            on the remaining max_num_batched_tokens.
830
        embedding_mode: Whether the running model is for embedding.
831
832
833
834
835
836
        preemption_mode: Whether to perform preemption by swapping or 
            recomputation. If not specified, we determine the mode as follows:
            We use recomputation by default since it incurs lower overhead than
            swapping. However, when the sequence group has multiple sequences
            (e.g., beam search), recomputation is not currently supported. In
            such a case, we use swapping instead.
837
    """
838

839
840
841
842
843
844
845
846
847
848
    def __init__(self,
                 max_num_batched_tokens: Optional[int],
                 max_num_seqs: int,
                 max_model_len: int,
                 use_v2_block_manager: bool = False,
                 num_lookahead_slots: int = 0,
                 delay_factor: float = 0.0,
                 enable_chunked_prefill: bool = False,
                 embedding_mode: Optional[bool] = False,
                 preemption_mode: Optional[str] = None) -> None:
849
850
851
        if max_num_batched_tokens is not None:
            self.max_num_batched_tokens = max_num_batched_tokens
        else:
852
            if enable_chunked_prefill:
853
854
855
                # It is the values that have the best balance between ITL
                # and TTFT on A100. Note it is not optimized for throughput.
                self.max_num_batched_tokens = 512
856
857
858
859
            elif embedding_mode:
                # For embedding, choose specific value for higher throughput
                self.max_num_batched_tokens = max(
                    max_model_len, _EMBEDDING_MODEL_MAX_NUM_BATCHED_TOKENS)
860
861
862
863
864
            else:
                # If max_model_len is too short, use 2048 as the default value
                # for higher throughput.
                self.max_num_batched_tokens = max(max_model_len, 2048)
        if enable_chunked_prefill:
865
866
            logger.info(
                "Chunked prefill is enabled with max_num_batched_tokens=%d.",
867
                self.max_num_batched_tokens)
868

869
        self.max_num_seqs = max_num_seqs
Lily Liu's avatar
Lily Liu committed
870
        self.max_model_len = max_model_len
871
        self.use_v2_block_manager = use_v2_block_manager
872
873
        self.num_lookahead_slots = num_lookahead_slots
        self.delay_factor = delay_factor
874
        self.chunked_prefill_enabled = enable_chunked_prefill
875
        self.embedding_mode = embedding_mode
876
        self.preemption_mode = preemption_mode
877
878
879
        self._verify_args()

    def _verify_args(self) -> None:
880
881
        if (self.max_num_batched_tokens < self.max_model_len
                and not self.chunked_prefill_enabled):
882
883
884
885
886
887
888
            raise ValueError(
                f"max_num_batched_tokens ({self.max_num_batched_tokens}) is "
                f"smaller than max_model_len ({self.max_model_len}). "
                "This effectively limits the maximum sequence length to "
                "max_num_batched_tokens and makes vLLM reject longer "
                "sequences. Please increase max_num_batched_tokens or "
                "decrease max_model_len.")
889

890
891
892
893
894
        if self.max_num_batched_tokens < self.max_num_seqs:
            raise ValueError(
                f"max_num_batched_tokens ({self.max_num_batched_tokens}) must "
                "be greater than or equal to max_num_seqs "
                f"({self.max_num_seqs}).")
895

896
897
898
899
900
901
        if self.num_lookahead_slots < 0:
            raise ValueError(
                "num_lookahead_slots "
                f"({self.num_lookahead_slots}) must be greater than or "
                "equal to 0.")

902

903
class DeviceConfig:
904
    device: Optional[torch.device]
905

906
907
908
    def __init__(self, device: str = "auto") -> None:
        if device == "auto":
            # Automated device type detection
909
            if is_neuron():
910
                self.device_type = "neuron"
911
912
            elif is_openvino():
                self.device_type = "openvino"
913
914
            elif is_tpu():
                self.device_type = "tpu"
915
916
            elif is_cpu():
                self.device_type = "cpu"
917
918
            elif is_xpu():
                self.device_type = "xpu"
919
            else:
920
921
922
                # We don't call torch.cuda.is_available() here to
                # avoid initializing CUDA before workers are forked
                self.device_type = "cuda"
923
924
925
926
927
        else:
            # Device type is assigned explicitly
            self.device_type = device

        # Some device types require processing inputs on CPU
928
        if self.device_type in ["neuron", "openvino"]:
929
            self.device = torch.device("cpu")
930
931
        elif self.device_type in ["tpu"]:
            self.device = None
932
933
934
935
        else:
            # Set device with device type
            self.device = torch.device(self.device_type)

936

937
938
939
940
941
942
943
944
945
946
947
948
949
class SpeculativeConfig:
    """Configuration for speculative decoding.

    The configuration is currently specialized to draft-model speculative
    decoding with top-1 proposals.
    """

    @staticmethod
    def maybe_create_spec_config(
        target_model_config: ModelConfig,
        target_parallel_config: ParallelConfig,
        target_dtype: str,
        speculative_model: Optional[str],
950
        speculative_draft_tensor_parallel_size: Optional[int],
951
        num_speculative_tokens: Optional[int],
952
953
954
        speculative_max_model_len: Optional[int],
        enable_chunked_prefill: bool,
        use_v2_block_manager: bool,
955
        disable_log_stats: bool,
956
        speculative_disable_by_batch_size: Optional[int],
957
958
        ngram_prompt_lookup_max: Optional[int],
        ngram_prompt_lookup_min: Optional[int],
959
960
961
        draft_token_acceptance_method: str,
        typical_acceptance_sampler_posterior_threshold: Optional[float],
        typical_acceptance_sampler_posterior_alpha: Optional[float],
962
        disable_logprobs: Optional[bool],
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
    ) -> Optional["SpeculativeConfig"]:
        """Create a SpeculativeConfig if possible, else return None.

        This function attempts to create a SpeculativeConfig object based on the
        provided parameters. If the necessary conditions are met, it returns an
        instance of SpeculativeConfig. Otherwise, it returns None.

        Args:
            target_model_config (ModelConfig): The configuration of the target
                model.
            target_parallel_config (ParallelConfig): The parallel configuration
                for the target model.
            target_dtype (str): The data type used for the target model.
            speculative_model (Optional[str]): The name of the speculative
                model, if provided.
978
979
            speculative_draft_tensor_parallel_size (Optional[int]): The degree
                of the tensor parallelism for the draft model.
980
            num_speculative_tokens (Optional[int]): The number of speculative
981
982
                tokens, if provided. Will default to the number in the draft
                model config if present, otherwise is required.
983
984
985
986
987
988
989
990
991
            speculative_max_model_len (Optional[int]): The maximum model len of
                the speculative model. Used when testing the ability to skip
                speculation for some sequences.
            enable_chunked_prefill (bool): Whether vLLM is configured to use
                chunked prefill or not. Used for raising an error since its not
                yet compatible with spec decode.
            use_v2_block_manager (bool): Whether vLLM is configured to use the
                v2 block manager or not. Used for raising an error since the v2
                block manager is required with spec decode.
992
993
994
            speculative_disable_by_batch_size (Optional[int]): Disable
                speculative decoding for new incoming requests when the number
                of enqueue requests  is larger than this value, if provided.
995
996
997
998
            ngram_prompt_lookup_max (Optional[int]): Max size of ngram token
                window, if provided.
            ngram_prompt_lookup_min (Optional[int]): Min size of ngram token
                window, if provided.
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
            draft_token_acceptance_method (str): The method to use for
                accepting draft tokens. This can take two possible
                values 'rejection_sampler' and 'typical_acceptance_sampler'
                for RejectionSampler and TypicalAcceptanceSampler
                respectively.
            typical_acceptance_sampler_posterior_threshold (Optional[float]):
                A threshold value that sets a lower bound on the posterior
                probability of a token in the target model for it to be
                accepted. This threshold is used only when we use the 
                TypicalAcceptanceSampler for token acceptance.
            typical_acceptance_sampler_posterior_alpha (Optional[float]):
                A scaling factor for the entropy-based threshold in the
                TypicalAcceptanceSampler.
1012
1013
1014
1015
1016
            disable_logprobs (Optional[bool]): If set to True, token log
                probabilities are not returned during speculative decoding.
                If set to False, token log probabilities are returned
                according to the log probability settings in SamplingParams.
                If not specified, it defaults to True.
1017
    
1018
1019
1020
1021
1022
        Returns:
            Optional["SpeculativeConfig"]: An instance of SpeculativeConfig if
                the necessary conditions are met, else None.
        """

1023
1024
1025
1026
        if speculative_model is None:
            if num_speculative_tokens is not None:
                raise ValueError("num_speculative_tokens was provided without "
                                 "speculative_model.")
1027
1028
            return None

1029
1030
1031
1032
1033
1034
        if (speculative_disable_by_batch_size is not None
                and speculative_disable_by_batch_size < 2):
            raise ValueError("Expect the batch size threshold of disabling "
                             "speculative decoding is > 1, but got "
                             f"{speculative_disable_by_batch_size=}")

1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
        if enable_chunked_prefill:
            raise ValueError(
                "Speculative decoding and chunked prefill are "
                f"currently mutually exclusive ({enable_chunked_prefill=}).")

        if not use_v2_block_manager:
            raise ValueError(
                "Speculative decoding requires usage of the V2 "
                "block manager. Enable it with --use-v2-block-manager.")

1045
1046
1047
1048
1049
1050
        # TODO: The user should be able to specify revision/quantization/max
        # model len for the draft model. It is not currently supported.
        draft_revision = None
        draft_code_revision = None
        draft_quantization = None

1051
1052
        if speculative_model == "[ngram]":
            if ngram_prompt_lookup_min is None:
1053
1054
1055
1056
1057
1058
1059
1060
                ngram_prompt_lookup_min = 1
            if ngram_prompt_lookup_max is None or ngram_prompt_lookup_max < 1:
                raise ValueError(f"{ngram_prompt_lookup_max=} must be > 0")
            if ngram_prompt_lookup_min < 1:
                raise ValueError(f"{ngram_prompt_lookup_min=} must be > 0")
            if ngram_prompt_lookup_min > ngram_prompt_lookup_max:
                raise ValueError(f"{ngram_prompt_lookup_min=} cannot be "
                                 f"larger than {ngram_prompt_lookup_max=}")
1061

1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
            # 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.
            draft_model_config = target_model_config
            draft_parallel_config = target_parallel_config
        else:
            ngram_prompt_lookup_max = 0
            ngram_prompt_lookup_min = 0
            draft_model_config = ModelConfig(
                model=speculative_model,
                tokenizer=target_model_config.tokenizer,
                tokenizer_mode=target_model_config.tokenizer_mode,
                trust_remote_code=target_model_config.trust_remote_code,
                dtype=target_model_config.dtype,
                seed=target_model_config.seed,
                revision=draft_revision,
                code_revision=draft_code_revision,
                tokenizer_revision=target_model_config.tokenizer_revision,
                max_model_len=None,
                quantization=draft_quantization,
                enforce_eager=target_model_config.enforce_eager,
1083
1084
                max_seq_len_to_capture=target_model_config.
                max_seq_len_to_capture,
1085
1086
1087
                max_logprobs=target_model_config.max_logprobs,
            )

1088
            draft_hf_config = draft_model_config.hf_config
1089

1090
1091
1092
1093
1094
            if (num_speculative_tokens is not None
                    and hasattr(draft_hf_config, "num_lookahead_tokens")):
                draft_hf_config.num_lookahead_tokens = num_speculative_tokens

            n_predict = getattr(draft_hf_config, "n_predict", None)
1095
1096
1097
1098
1099
1100
1101
1102
            if n_predict is not None:
                if num_speculative_tokens is None:
                    # Default to max value defined in draft model config.
                    num_speculative_tokens = n_predict
                elif num_speculative_tokens > n_predict:
                    # Verify provided value doesn't exceed the maximum
                    # supported by the draft model.
                    raise ValueError(
1103
1104
1105
                        "This speculative model supports a maximum of "
                        f"num_speculative_tokens={n_predict}, but "
                        f"{num_speculative_tokens=} was provided.")
1106

1107
1108
1109
1110
1111
1112
1113
1114
1115
            draft_model_config.max_model_len = (
                SpeculativeConfig._maybe_override_draft_max_model_len(
                    speculative_max_model_len,
                    draft_model_config.max_model_len,
                    target_model_config.max_model_len,
                ))

            draft_parallel_config = (
                SpeculativeConfig.create_draft_parallel_config(
1116
                    target_parallel_config,
1117
                    speculative_draft_tensor_parallel_size, draft_hf_config))
1118

1119
1120
1121
1122
1123
1124
        if num_speculative_tokens is None:
            raise ValueError(
                "num_speculative_tokens must be provided with "
                "speculative_model unless the draft model config contains an "
                "n_predict parameter.")

1125
1126
1127
1128
        if typical_acceptance_sampler_posterior_threshold is None:
            typical_acceptance_sampler_posterior_threshold = 0.09
        if typical_acceptance_sampler_posterior_alpha is None:
            typical_acceptance_sampler_posterior_alpha = 0.3
1129
1130
        if disable_logprobs is None:
            disable_logprobs = True
1131

1132
1133
1134
1135
        return SpeculativeConfig(
            draft_model_config,
            draft_parallel_config,
            num_speculative_tokens,
1136
            speculative_disable_by_batch_size,
1137
1138
            ngram_prompt_lookup_max,
            ngram_prompt_lookup_min,
1139
1140
1141
1142
1143
            draft_token_acceptance_method=draft_token_acceptance_method,
            typical_acceptance_sampler_posterior_threshold=\
                typical_acceptance_sampler_posterior_threshold,
            typical_acceptance_sampler_posterior_alpha=\
                typical_acceptance_sampler_posterior_alpha,
1144
1145
            disable_logprobs=disable_logprobs,
            disable_log_stats=disable_log_stats,
1146
1147
        )

1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
    @staticmethod
    def _maybe_override_draft_max_model_len(
        speculative_max_model_len: Optional[int],
        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:
                raise ValueError(f"{speculative_max_model_len=} cannot be "
                                 f"larger than {draft_max_model_len=}")

            if speculative_max_model_len > target_max_model_len:
                raise ValueError(f"{speculative_max_model_len=} cannot be "
                                 f"larger than {target_max_model_len=}")

            return speculative_max_model_len

        return min(
            draft_max_model_len,
            target_max_model_len,
        )

1183
1184
    @staticmethod
    def create_draft_parallel_config(
1185
        target_parallel_config: ParallelConfig,
1186
1187
        speculative_draft_tensor_parallel_size: Optional[int],
        draft_hf_config: PretrainedConfig,
1188
    ) -> ParallelConfig:
1189
1190
        """Create a parallel config for use by the draft worker.

1191
        This is mostly a copy of the target parallel config, except the tp_size.
1192
        """
1193
        if speculative_draft_tensor_parallel_size is None:
1194
1195
1196
1197
1198
1199
1200
1201
1202
            if draft_hf_config.model_type == "mlp_speculator":
                speculative_draft_tensor_parallel_size = 1
                if target_parallel_config.tensor_parallel_size > 1:
                    logger.warning(
                        "MLPSpeculator cannot currently be run with tp>1; "
                        "setting speculative_draft_tensor_parallel_size=1")
            else:
                speculative_draft_tensor_parallel_size = \
                    target_parallel_config.tensor_parallel_size
1203
1204
1205
1206
1207
1208
        elif speculative_draft_tensor_parallel_size != 1:
            # TODO(wooyeon): allow tp values larger than 1
            raise ValueError(
                f"{speculative_draft_tensor_parallel_size=} cannot be"
                f"other value than 1")

1209
1210
1211
        draft_parallel_config = ParallelConfig(
            pipeline_parallel_size=target_parallel_config.
            pipeline_parallel_size,
1212
            tensor_parallel_size=speculative_draft_tensor_parallel_size,
1213
1214
            distributed_executor_backend=target_parallel_config.
            distributed_executor_backend,
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
            max_parallel_loading_workers=target_parallel_config.
            max_parallel_loading_workers,
            disable_custom_all_reduce=target_parallel_config.
            disable_custom_all_reduce,
            tokenizer_pool_config=target_parallel_config.tokenizer_pool_config,
            ray_workers_use_nsight=target_parallel_config.
            ray_workers_use_nsight,
            placement_group=target_parallel_config.placement_group,
        )

        return draft_parallel_config

    def __init__(
        self,
        draft_model_config: ModelConfig,
        draft_parallel_config: ParallelConfig,
        num_speculative_tokens: int,
1232
1233
1234
        speculative_disable_by_batch_size: Optional[int],
        ngram_prompt_lookup_max: Optional[int],
        ngram_prompt_lookup_min: Optional[int],
1235
1236
1237
        draft_token_acceptance_method: str,
        typical_acceptance_sampler_posterior_threshold: float,
        typical_acceptance_sampler_posterior_alpha: float,
1238
        disable_logprobs: bool,
1239
        disable_log_stats: bool,
1240
1241
1242
1243
1244
1245
1246
1247
    ):
        """Create a SpeculativeConfig object.

        Args:
            draft_model_config: ModelConfig for the draft model.
            draft_parallel_config: ParallelConfig for the draft model.
            num_speculative_tokens: The number of tokens to sample from the
                draft model before scoring with the target model.
1248
1249
1250
1251
1252
            speculative_disable_by_batch_size: Disable speculative
                decoding for new incoming requests when the number of
                enqueue requests is larger than this value.
            ngram_prompt_lookup_max: Max size of ngram token window.
            ngram_prompt_lookup_min: Min size of ngram token window.
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
            draft_token_acceptance_method (str): The method to use for
                accepting draft tokens. This can take two possible
                values 'rejection_sampler' and 'typical_acceptance_sampler'
                for RejectionSampler and TypicalAcceptanceSampler
                respectively.
            typical_acceptance_sampler_posterior_threshold (Optional[float]):
                A threshold value that sets a lower bound on the posterior
                probability of a token in the target model for it to be
                accepted. This threshold is used only when we use the 
                TypicalAcceptanceSampler for token acceptance.
            typical_acceptance_sampler_posterior_alpha (Optional[float]):
                A scaling factor for the entropy-based threshold in the
                TypicalAcceptanceSampler.
1266
1267
1268
1269
1270
1271
            disable_logprobs: If set to True, token log probabilities will not
                be returned even if requested by sampling parameters. This 
                reduces latency by skipping logprob calculation in proposal
                sampling, target sampling, and after accepted tokens are
                determined. If set to False, log probabilities will be
                returned.
1272
1273
            disable_log_stats: Whether to disable periodic printing of stage
                times in speculative decoding.
1274
1275
1276
1277
        """
        self.draft_model_config = draft_model_config
        self.draft_parallel_config = draft_parallel_config
        self.num_speculative_tokens = num_speculative_tokens
1278
1279
1280
1281
        self.speculative_disable_by_batch_size = \
            speculative_disable_by_batch_size
        self.ngram_prompt_lookup_max = ngram_prompt_lookup_max or 0
        self.ngram_prompt_lookup_min = ngram_prompt_lookup_min or 0
1282
1283
1284
1285
1286
        self.draft_token_acceptance_method = draft_token_acceptance_method
        self.typical_acceptance_sampler_posterior_threshold = \
            typical_acceptance_sampler_posterior_threshold
        self.typical_acceptance_sampler_posterior_alpha = \
            typical_acceptance_sampler_posterior_alpha
1287
        self.disable_logprobs = disable_logprobs
1288
        self.disable_log_stats = disable_log_stats
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299

        self._verify_args()

    def _verify_args(self) -> None:
        if self.num_speculative_tokens <= 0:
            raise ValueError("Expected num_speculative_tokens to be greater "
                             f"than zero ({self.num_speculative_tokens}).")

        if self.draft_model_config:
            self.draft_model_config.verify_with_parallel_config(
                self.draft_parallel_config)
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
            # Validate and set draft token acceptance related settings.

        if (self.draft_token_acceptance_method is None):
            raise ValueError("draft_token_acceptance_method is not set. "
                             "Expected values are rejection_sampler or "
                             "typical_acceptance_sampler.")

        if (self.draft_token_acceptance_method != 'rejection_sampler'
                and self.draft_token_acceptance_method !=
                'typical_acceptance_sampler'):
            raise ValueError(
                "Expected draft_token_acceptance_method to be either "
                "rejection_sampler or typical_acceptance_sampler. Instead it "
                f"is {self.draft_token_acceptance_method}")

        if (self.typical_acceptance_sampler_posterior_threshold < 0
                or self.typical_acceptance_sampler_posterior_alpha < 0):
            raise ValueError(
                "Expected typical_acceptance_sampler_posterior_threshold "
                "and typical_acceptance_sampler_posterior_alpha to be > 0. "
                "Instead found "
                f"typical_acceptance_sampler_posterior_threshold = "
                f"{self.typical_acceptance_sampler_posterior_threshold} and "
                f"typical_acceptance_sampler_posterior_alpha = "
                f"{self.typical_acceptance_sampler_posterior_alpha}")
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336

    @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 __repr__(self) -> str:
1337
1338
1339
1340
        if self.ngram_prompt_lookup_max > 0:
            draft_model = "[ngram]"
        else:
            draft_model = self.draft_model_config.model
1341
1342
1343
1344
        num_spec_tokens = self.num_speculative_tokens
        return f"SpeculativeConfig({draft_model=}, {num_spec_tokens=})"


1345
1346
1347
1348
@dataclass
class LoRAConfig:
    max_lora_rank: int
    max_loras: int
1349
    fully_sharded_loras: bool = False
1350
1351
1352
1353
1354
    max_cpu_loras: Optional[int] = None
    lora_dtype: Optional[torch.dtype] = None
    lora_extra_vocab_size: int = 256
    # This is a constant.
    lora_vocab_padding_size: ClassVar[int] = 256
1355
    long_lora_scaling_factors: Optional[Tuple[float]] = None
1356
1357

    def __post_init__(self):
1358
1359
1360
        # Setting the maximum rank to 256 should be able to satisfy the vast
        # majority of applications.
        possible_max_ranks = (8, 16, 32, 64, 128, 256)
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
        possible_lora_extra_vocab_size = (0, 256, 512)
        if self.max_lora_rank not in possible_max_ranks:
            raise ValueError(
                f"max_lora_rank ({self.max_lora_rank}) must be one of "
                f"{possible_max_ranks}.")
        if self.lora_extra_vocab_size not in possible_lora_extra_vocab_size:
            raise ValueError(
                f"lora_extra_vocab_size ({self.lora_extra_vocab_size}) "
                f"must be one of {possible_lora_extra_vocab_size}.")
        if self.max_loras < 1:
            raise ValueError(f"max_loras ({self.max_loras}) must be >= 1.")
        if self.max_cpu_loras is None:
            self.max_cpu_loras = self.max_loras
        elif self.max_cpu_loras < self.max_loras:
            raise ValueError(
                f"max_cpu_loras ({self.max_cpu_loras}) must be >= "
zspo's avatar
zspo committed
1377
                f"max_loras ({self.max_loras})")
1378
1379
1380
1381
1382
1383

    def verify_with_model_config(self, model_config: ModelConfig):
        if self.lora_dtype in (None, "auto"):
            self.lora_dtype = model_config.dtype
        elif isinstance(self.lora_dtype, str):
            self.lora_dtype = getattr(torch, self.lora_dtype)
1384
1385
1386
1387
        if model_config.quantization and model_config.quantization not in [
                "awq", "gptq"
        ]:
            # TODO support marlin and squeezellm
1388
1389
            logger.warning("%s quantization is not tested with LoRA yet.",
                           model_config.quantization)
1390
1391

    def verify_with_scheduler_config(self, scheduler_config: SchedulerConfig):
1392
1393
        if scheduler_config.chunked_prefill_enabled:
            raise ValueError("LoRA is not supported with chunked prefill yet.")
1394
1395


1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
@dataclass
class PromptAdapterConfig:
    max_prompt_adapters: int
    max_prompt_adapter_token: int
    max_cpu_prompt_adapters: Optional[int] = None
    prompt_adapter_dtype: Optional[torch.dtype] = None

    def __post_init__(self):
        library_name = 'peft'
        try:
            __import__(library_name)
        except ImportError as e:
            raise ImportError(
                f"'{library_name}' is not installed for prompt adapter support."
                f"Please install it using 'pip install {library_name}'."
            ) from e

        if self.max_prompt_adapters < 1:
            raise ValueError(f"max_prompt_adapters "
                             f"({self.max_prompt_adapters}) must be >= 1.")
        if self.max_prompt_adapter_token == 0:
            raise ValueError("max_prompt_adapter_token must be set.")
        if self.max_cpu_prompt_adapters is None:
            self.max_cpu_prompt_adapters = self.max_prompt_adapters

    def verify_with_model_config(self, model_config: ModelConfig):
        if self.prompt_adapter_dtype in (None, "auto"):
            self.prompt_adapter_dtype = model_config.dtype
        elif isinstance(self.prompt_adapter_dtype, str):
            self.prompt_adapter_dtype = getattr(torch,
                                                self.prompt_adapter_dtype)


1429
@dataclass
1430
class MultiModalConfig:
1431
    """Configs the input data format and how models should run for
1432
1433
1434
    multimodal models."""
    # TODO: Add configs to init vision tower or not.
    pass
1435

1436

1437
1438
1439
1440
1441
1442
1443
1444
_STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.float16,
    "float16": torch.float16,
    "float": torch.float32,
    "float32": torch.float32,
    "bfloat16": torch.bfloat16,
}

1445
_ROCM_NOT_SUPPORTED_DTYPE: List[str] = []  #
1446

1447
1448
1449

def _get_and_verify_dtype(
    config: PretrainedConfig,
1450
    dtype: Union[str, torch.dtype],
1451
1452
1453
1454
1455
1456
1457
) -> torch.dtype:
    # NOTE: getattr(config, "torch_dtype", torch.float32) is not correct
    # because config.torch_dtype can be None.
    config_dtype = getattr(config, "torch_dtype", None)
    if config_dtype is None:
        config_dtype = torch.float32

1458
1459
1460
1461
    if isinstance(dtype, str):
        dtype = dtype.lower()
        if dtype == "auto":
            if config_dtype == torch.float32:
Woosuk Kwon's avatar
Woosuk Kwon committed
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
                if config.model_type == "gemma2":
                    logger.info(
                        "For Gemma 2, we downcast float32 to bfloat16 instead "
                        "of float16 by default. Please specify `dtype` if you "
                        "want to use float16.")
                    torch_dtype = torch.bfloat16
                else:
                    # Following the common practice, we use float16 for float32
                    # models.
                    torch_dtype = torch.float16
1472
1473
            else:
                torch_dtype = config_dtype
1474
        else:
1475
1476
1477
1478
1479
            if dtype not in _STR_DTYPE_TO_TORCH_DTYPE:
                raise ValueError(f"Unknown dtype: {dtype}")
            torch_dtype = _STR_DTYPE_TO_TORCH_DTYPE[dtype]
    elif isinstance(dtype, torch.dtype):
        torch_dtype = dtype
1480
    else:
1481
        raise ValueError(f"Unknown dtype: {dtype}")
1482
1483
1484
1485
1486

    # Verify the dtype.
    if torch_dtype != config_dtype:
        if torch_dtype == torch.float32:
            # Upcasting to float32 is allowed.
1487
            logger.info("Upcasting %s to %s.", config_dtype, torch_dtype)
1488
1489
1490
            pass
        elif config_dtype == torch.float32:
            # Downcasting from float32 to float16 or bfloat16 is allowed.
1491
            logger.info("Downcasting %s to %s.", config_dtype, torch_dtype)
1492
1493
            pass
        else:
Woosuk Kwon's avatar
Woosuk Kwon committed
1494
            # Casting between float16 and bfloat16 is allowed with a warning.
1495
            logger.warning("Casting %s to %s.", config_dtype, torch_dtype)
1496
1497

    return torch_dtype
1498
1499
1500
1501
1502


def _get_and_verify_max_len(
    hf_config: PretrainedConfig,
    max_model_len: Optional[int],
1503
1504
    disable_sliding_window: bool,
    sliding_window_len: Optional[int],
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
) -> int:
    """Get and verify the model's maximum length."""
    derived_max_model_len = float("inf")
    possible_keys = [
        # OPT
        "max_position_embeddings",
        # GPT-2
        "n_positions",
        # MPT
        "max_seq_len",
1515
1516
        # ChatGLM2
        "seq_length",
1517
1518
        # Command-R
        "model_max_length",
1519
1520
1521
1522
1523
        # Others
        "max_sequence_length",
        "max_seq_length",
        "seq_len",
    ]
1524
    # Choose the smallest "max_length" from the possible keys.
1525
    max_len_key = None
1526
    for key in possible_keys:
1527
1528
1529
1530
1531
        max_len = getattr(hf_config, key, None)
        if max_len is not None:
            max_len_key = key if max_len < derived_max_model_len \
                else max_len_key
            derived_max_model_len = min(derived_max_model_len, max_len)
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541

    # If sliding window is manually disabled, max_length should be less
    # than the sliding window length in the model config.
    if disable_sliding_window and sliding_window_len is not None:
        max_len_key = "sliding_window" \
            if sliding_window_len < derived_max_model_len else max_len_key
        derived_max_model_len = min(derived_max_model_len, sliding_window_len)

    # If none of the keys were found in the config, use a default and
    # log a warning.
1542
    if derived_max_model_len == float("inf"):
1543
1544
1545
1546
1547
1548
1549
1550
        if max_model_len is not None:
            # If max_model_len is specified, we use it.
            return max_model_len

        default_max_len = 2048
        logger.warning(
            "The model's config.json does not contain any of the following "
            "keys to determine the original maximum length of the model: "
1551
            "%s. Assuming the model's maximum length is %d.", possible_keys,
1552
            default_max_len)
1553
        derived_max_model_len = default_max_len
1554

1555
    rope_scaling = getattr(hf_config, "rope_scaling", None)
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
    if rope_scaling is not None:
        if "type" in rope_scaling:
            rope_type = rope_scaling["type"]
        elif "rope_type" in rope_scaling:
            rope_type = rope_scaling["rope_type"]
        else:
            raise ValueError(
                "rope_scaling must have a 'type' or 'rope_type' key.")

        # The correct one should be "longrope", kept "su" here
        # to be backward compatible
        if rope_type not in ("su", "longrope", "llama3"):
            if disable_sliding_window:
                # TODO(robertgshaw): Find a model that supports rope_scaling
                # with sliding window to see if this case should be allowed.
                raise NotImplementedError(
                    "Disabling sliding window is not supported for models "
                    "with rope_scaling. Please raise an issue so we can "
                    "investigate.")

            assert "factor" in rope_scaling
            scaling_factor = rope_scaling["factor"]
            if rope_type == "yarn":
                derived_max_model_len = rope_scaling[
                    "original_max_position_embeddings"]
            derived_max_model_len *= scaling_factor
1582

1583
1584
    # If the user specified a max length, make sure it is smaller than the
    # derived length from the HF model config.
1585
    if max_model_len is None:
1586
        max_model_len = int(derived_max_model_len)
1587
    elif max_model_len > derived_max_model_len:
1588
1589
1590
1591
1592
        # Some models might have a separate key for specifying model_max_length
        # that will be bigger than derived_max_model_len. We compare user input
        # with model_max_length and allow this override when it's smaller.
        model_max_length = getattr(hf_config, "model_max_length", None)
        if model_max_length is not None and max_model_len <= model_max_length:
1593
1594
1595
1596
1597
1598
1599
            if disable_sliding_window:
                # TODO(robertgshaw): Find a model that has model_max_length
                # with sliding window to see if this case should be allowed.
                raise NotImplementedError(
                    "Disabling sliding window is not supported for models "
                    "model_max_length in the config. Please raise an issue "
                    "so we can investigate.")
1600
        else:
1601
            msg = (
1602
                f"User-specified max_model_len ({max_model_len}) is greater "
1603
1604
                f"than the derived max_model_len ({max_len_key}="
                f"{derived_max_model_len} or model_max_length="
1605
                f"{model_max_length} in model's config.json). This may lead "
1606
1607
1608
1609
1610
1611
1612
1613
1614
                "to incorrect model outputs or CUDA errors.")
            if envs.VLLM_ALLOW_LONG_MAX_MODEL_LEN:
                logger.warning(
                    "%s Make sure the value is correct and within the "
                    "model context size.", msg)
            else:
                raise ValueError(
                    f"{msg} To allow overriding this maximum, set "
                    "the env var VLLM_ALLOW_LONG_MAX_MODEL_LEN=1")
1615
    return int(max_model_len)
1616
1617


1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
def get_served_model_name(model: str,
                          served_model_name: Optional[Union[str, List[str]]]):
    """
    If the input is a non-empty list, the first model_name in 
    `served_model_name` is taken. 
    If the input is a non-empty string, it is used directly. 
    For cases where the input is either an empty string or an 
    empty list, the fallback is to use `self.model`.
    """
    if not served_model_name:
        return model
    if isinstance(served_model_name, list):
        return served_model_name[0]
    return served_model_name


1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
@dataclass
class DecodingConfig:
    """Dataclass which contains the decoding strategy of the engine"""

    # Which guided decoding algo to use. 'outlines' / 'lm-format-enforcer'
    guided_decoding_backend: str = 'outlines'

    def __post_init__(self):
        valid_guided_backends = ['outlines', 'lm-format-enforcer']
        backend = self.guided_decoding_backend
        if backend not in valid_guided_backends:
            raise ValueError(f"Invalid guided_decoding_backend '{backend},"
                             f"must be one of {valid_guided_backends}")


1649
1650
1651
1652
1653
@dataclass
class ObservabilityConfig:
    """Configuration for observability."""
    otlp_traces_endpoint: Optional[str] = None

1654
1655
1656
1657
1658
1659
1660
1661
    # Collecting detailed timing information for each request can be expensive.

    # If set, collects the model forward time for the request.
    collect_model_forward_time: bool = False

    # If set, collects the model execute time for the request.
    collect_model_execute_time: bool = False

1662
1663
1664
1665
1666
    def __post_init__(self):
        if not is_otel_installed() and self.otlp_traces_endpoint is not None:
            raise ValueError("OpenTelemetry packages must be installed before "
                             "configuring 'otlp_traces_endpoint'")

1667
1668
1669
1670
1671
1672
1673
        if ((self.collect_model_forward_time
             or self.collect_model_execute_time)
                and self.otlp_traces_endpoint is None):
            raise ValueError(
                "collect_model_forward_time or collect_model_execute_time "
                "requires --otlp-traces-endpoint to be set.")

1674

1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
@dataclass(frozen=True)
class EngineConfig:
    """Dataclass which contains all engine-related configuration. This
    simplifies passing around the distinct configurations in the codebase.
    """

    model_config: ModelConfig
    cache_config: CacheConfig
    parallel_config: ParallelConfig
    scheduler_config: SchedulerConfig
    device_config: DeviceConfig
1686
    load_config: LoadConfig
1687
    lora_config: Optional[LoRAConfig]
1688
    multimodal_config: Optional[MultiModalConfig]
1689
    speculative_config: Optional[SpeculativeConfig]
1690
    decoding_config: Optional[DecodingConfig]
1691
    observability_config: Optional[ObservabilityConfig]
1692
    prompt_adapter_config: Optional[PromptAdapterConfig]
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703

    def __post_init__(self):
        """Verify configs are valid & consistent with each other.
        """
        self.model_config.verify_with_parallel_config(self.parallel_config)
        self.cache_config.verify_with_parallel_config(self.parallel_config)

        if self.lora_config:
            self.lora_config.verify_with_model_config(self.model_config)
            self.lora_config.verify_with_scheduler_config(
                self.scheduler_config)
1704
1705
1706
        if self.prompt_adapter_config:
            self.prompt_adapter_config.verify_with_model_config(
                self.model_config)
1707
1708
1709
1710
1711
1712

    def to_dict(self):
        """Return the configs as a dictionary, for use in **kwargs.
        """
        return dict(
            (field.name, getattr(self, field.name)) for field in fields(self))