config.py 71.3 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

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

18
19
20
if TYPE_CHECKING:
    from ray.util.placement_group import PlacementGroup

21
    from vllm.executor.executor_base import ExecutorBase
22
    from vllm.model_executor.model_loader.loader import BaseModelLoader
23
24
    from vllm.transformers_utils.tokenizer_group.base_tokenizer_group import (
        BaseTokenizerGroup)
25

26
27
logger = init_logger(__name__)

28
_GB = 1 << 30
29
_EMBEDDING_MODEL_MAX_NUM_BATCHED_TOKENS = 32768
30

31
32
33
_PP_SUPPORTED_MODELS = [
    "AquilaModel",
    "AquilaForCausalLM",
34
    "DeepseekV2ForCausalLM",
35
36
37
38
39
40
    "InternLMForCausalLM",
    "LlamaForCausalLM",
    "LLaMAForCausalLM",
    "MistralForCausalLM",
    "Phi3ForCausalLM",
    "GPT2LMHeadModel",
41
    "MixtralForCausalLM",
42
    "NemotronForCausalLM",
43
44
    "Qwen2ForCausalLM",
    "Qwen2MoeForCausalLM",
45
    "QWenLMHeadModel",
46
47
]

48
49

class ModelConfig:
50
51
52
53
    """Configuration for the model.

    Args:
        model: Name or path of the huggingface model to use.
54
55
            It is also used as the content for `model_name` tag in metrics 
            output when `served_model_name` is not specified. 
56
        tokenizer: Name or path of the huggingface tokenizer to use.
57
58
        tokenizer_mode: Tokenizer mode. "auto" will use the fast tokenizer if
            available, and "slow" will always use the slow tokenizer.
59
60
        trust_remote_code: Trust remote code (e.g., from HuggingFace) when
            downloading the model and tokenizer.
61
62
63
64
        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
65
66
67
        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.
68
        code_revision: The specific revision to use for the model code on
69
            Hugging Face Hub. It can be a branch name, a tag name, or a
70
            commit id. If unspecified, will use the default version.
71
72
73
        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.
74
75
76
        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.
77
78
        max_model_len: Maximum length of a sequence (including prompt and
            output). If None, will be derived from the model.
79
80
        quantization: Quantization method that was used to quantize the model
            weights. If None, we assume the model weights are not quantized.
81
82
        quantization_param_path: Path to JSON file containing scaling factors.
            Used to load KV cache scaling factors into the model when KV cache
83
84
            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
85
            model dtype is FP8_E4M3 on ROCm.
86
87
88
89
90
        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.
        max_context_len_to_capture: Maximum context len covered by CUDA graphs.
            When a sequence has context length larger than this, we fall back
91
92
93
94
            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
95
96
97
98
        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.
99
100
        skip_tokenizer_init: If true, skip initialization of tokenizer and
            detokenizer.
101
102
103
104
        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`.
105
    """
106
107
108
109

    def __init__(
        self,
        model: str,
110
111
        tokenizer: str,
        tokenizer_mode: str,
112
        trust_remote_code: bool,
113
        dtype: Union[str, torch.dtype],
114
        seed: int,
115
        revision: Optional[str] = None,
116
        code_revision: Optional[str] = None,
117
        rope_scaling: Optional[dict] = None,
118
        rope_theta: Optional[float] = None,
119
        tokenizer_revision: Optional[str] = None,
120
        max_model_len: Optional[int] = None,
121
        quantization: Optional[str] = None,
122
        quantization_param_path: Optional[str] = None,
123
124
        enforce_eager: bool = False,
        max_context_len_to_capture: Optional[int] = None,
125
        max_seq_len_to_capture: Optional[int] = None,
126
        max_logprobs: int = 20,
127
        disable_sliding_window: bool = False,
128
        skip_tokenizer_init: bool = False,
129
        served_model_name: Optional[Union[str, List[str]]] = None,
130
        multimodal_config: Optional["MultiModalConfig"] = None,
131
132
    ) -> None:
        self.model = model
133
        self.tokenizer = tokenizer
134
        self.tokenizer_mode = tokenizer_mode
135
        self.trust_remote_code = trust_remote_code
136
        self.seed = seed
Jasmond L's avatar
Jasmond L committed
137
        self.revision = revision
138
        self.code_revision = code_revision
139
        self.rope_scaling = rope_scaling
140
        self.rope_theta = rope_theta
141
142
143
144
145
        # 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
146
        self.quantization = quantization
147
        self.quantization_param_path = quantization_param_path
148
        self.enforce_eager = enforce_eager
149
        if max_context_len_to_capture is not None:
150
151
            raise ValueError("`max_context_len_to_capture` is deprecated. "
                             "Use `max_seq_len_to_capture` instead.")
152
        self.max_seq_len_to_capture = max_seq_len_to_capture
153
        self.max_logprobs = max_logprobs
154
        self.disable_sliding_window = disable_sliding_window
155
        self.skip_tokenizer_init = skip_tokenizer_init
156

157
        self.hf_config = get_config(self.model, trust_remote_code, revision,
158
                                    code_revision, rope_scaling, rope_theta)
159
160
        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
161
162
163
164
165
166
167
168
169
170
171

        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

172
173
174
175
176
        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())
177
178
        self.served_model_name = get_served_model_name(model,
                                                       served_model_name)
179
180
        self.multimodal_config = multimodal_config

181
182
        if not self.skip_tokenizer_init:
            self._verify_tokenizer_mode()
183
        self._verify_embedding_mode()
184
        self._verify_quantization()
185
        self._verify_cuda_graph()
186
187
188
189
190
191
192
193

    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
194

195
196
197
198
199
    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)

200
201
202
    def _parse_quant_hf_config(self):
        quant_cfg = getattr(self.hf_config, "quantization_config", None)
        if quant_cfg is None:
203
            # compressed-tensors uses a "compression_config" key
204
            quant_cfg = getattr(self.hf_config, "compression_config", None)
205
206
        return quant_cfg

207
    def _verify_quantization(self) -> None:
208
209
        supported_quantization = [*QUANTIZATION_METHODS]
        rocm_supported_quantization = ["gptq", "squeezellm"]
210
211
212
213
        optimized_quantization_methods = [
            "fp8", "marlin", "gptq_marlin_24", "gptq_marlin", "awq_marlin",
            "fbgemm_fp8", "compressed_tensors", "compressed-tensors"
        ]
214
215
216
217
        if self.quantization is not None:
            self.quantization = self.quantization.lower()

        # Parse quantization method from the HF model config, if available.
218
219
        quant_cfg = self._parse_quant_hf_config()

220
221
        if quant_cfg is not None:
            quant_method = quant_cfg.get("quant_method", "").lower()
222
223

            # Detect which checkpoint is it
224
            for _, method in QUANTIZATION_METHODS.items():
225
226
227
228
229
230
                quantization_override = method.override_quantization_method(
                    quant_cfg, self.quantization)
                if quantization_override:
                    quant_method = quantization_override
                    self.quantization = quantization_override
                    break
231

232
            # Verify quantization configurations.
233
            if self.quantization is None:
234
235
                self.quantization = quant_method
            elif self.quantization != quant_method:
236
237
                raise ValueError(
                    "Quantization method specified in the model config "
238
                    f"({quant_method}) does not match the quantization "
239
240
241
242
243
244
245
246
                    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}.")
247
            if is_hip(
248
            ) and self.quantization not in rocm_supported_quantization:
249
                raise ValueError(
250
251
                    f"{self.quantization} quantization is currently not "
                    f"supported in ROCm.")
252
            if self.quantization not in optimized_quantization_methods:
253
                logger.warning(
254
                    "%s quantization is not fully "
255
                    "optimized yet. The speed can be slower than "
256
                    "non-quantized models.", self.quantization)
257

258
    def _verify_cuda_graph(self) -> None:
259
260
261
262
        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)
263

264
265
266
267
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
268
269
        total_num_attention_heads = getattr(self.hf_text_config,
                                            "num_attention_heads", 0)
270
271
272
273
274
275
276
277
        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
278
279
280
281
282
283
284
        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}.")

285
286
287
288
289
290
        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.")

291
292
293
294
        if self.quantization == "bitsandbytes" and self.enforce_eager is False:
            raise ValueError(
                "BitAndBytes with enforce_eager = False is not supported yet.")

295
    def get_hf_config_sliding_window(self) -> Optional[int]:
Woosuk Kwon's avatar
Woosuk Kwon committed
296
        """Get the sliding window size, or None if disabled."""
297
298
299
300

        # 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.
301
302
        if (hasattr(self.hf_text_config, "use_sliding_window")
                and not self.hf_text_config.use_sliding_window):
303
            return None
304
        return getattr(self.hf_text_config, "sliding_window", None)
305

306
307
308
309
310
311
312
313
314
    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()

315
    def get_vocab_size(self) -> int:
316
        return self.hf_text_config.vocab_size
317

318
    def get_hidden_size(self) -> int:
319
        return self.hf_text_config.hidden_size
320
321

    def get_head_size(self) -> int:
wangding zeng's avatar
wangding zeng committed
322
323
324
325
326
327
        # 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
328
329
        if hasattr(self.hf_text_config, "head_dim"):
            return self.hf_text_config.head_dim
330
        # FIXME(woosuk): This may not be true for all models.
331
332
        return (self.hf_text_config.hidden_size //
                self.hf_text_config.num_attention_heads)
333

334
335
    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
Zhuohan Li's avatar
Zhuohan Li committed
336
        # For GPTBigCode & Falcon:
337
        # NOTE: for falcon, when new_decoder_architecture is True, the
Zhuohan Li's avatar
Zhuohan Li committed
338
339
        # multi_query flag is ignored and we use n_head_kv for the number of
        # KV heads.
340
        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
341
        new_decoder_arch_falcon = (
342
            self.hf_config.model_type in falcon_model_types
343
            and getattr(self.hf_config, "new_decoder_architecture", False))
344
        if not new_decoder_arch_falcon and getattr(self.hf_text_config,
345
                                                   "multi_query", False):
Zhuohan Li's avatar
Zhuohan Li committed
346
            # Multi-query attention, only one KV head.
Woosuk Kwon's avatar
Woosuk Kwon committed
347
            # Currently, tensor parallelism is not supported in this case.
Zhuohan Li's avatar
Zhuohan Li committed
348
            return 1
349

350
        # For DBRX and MPT
351
352
353
354
355
        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":
356
357
358
            return getattr(self.hf_config.attn_config, "kv_n_heads",
                           self.hf_config.num_attention_heads)

359
360
361
362
363
364
365
366
367
368
        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:
369
            num_kv_heads = getattr(self.hf_text_config, attr, None)
370
371
372
373
374
            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.
375
        return self.hf_text_config.num_attention_heads
376
377
378
379
380
381
382
383
384
385

    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)
386

387
388
    def get_num_attention_heads(self,
                                parallel_config: "ParallelConfig") -> int:
389
390
        num_heads = getattr(self.hf_text_config, "num_attention_heads", 0)
        return num_heads // parallel_config.tensor_parallel_size
391

392
    def get_num_layers(self, parallel_config: "ParallelConfig") -> int:
393
        from vllm.distributed.utils import get_pp_indices
Mor Zusman's avatar
Mor Zusman committed
394
395
        total_num_hidden_layers = getattr(self.hf_text_config,
                                          "num_hidden_layers", 0)
396
397
398
399
        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
400

Mor Zusman's avatar
Mor Zusman committed
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
    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"
        ])

427
428

class CacheConfig:
429
430
431
432
433
    """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
434
            vLLM execution.
435
        swap_space: Size of the CPU swap space per GPU (in GiB).
436
        cache_dtype: Data type for kv cache storage.
437
        num_gpu_blocks_override: Number of GPU blocks to use. This overrides the
438
            profiled num_gpu_blocks if specified. Does nothing if None.
439
    """
440

441
442
443
444
445
    def __init__(
        self,
        block_size: int,
        gpu_memory_utilization: float,
        swap_space: int,
446
        cache_dtype: str,
447
        num_gpu_blocks_override: Optional[int] = None,
448
        sliding_window: Optional[int] = None,
449
        enable_prefix_caching: bool = False,
450
        cpu_offload_gb: float = 0,
451
452
453
    ) -> None:
        self.block_size = block_size
        self.gpu_memory_utilization = gpu_memory_utilization
454
        self.swap_space_bytes = swap_space * _GB
455
        self.num_gpu_blocks_override = num_gpu_blocks_override
456
        self.cache_dtype = cache_dtype
457
        self.sliding_window = sliding_window
458
        self.enable_prefix_caching = enable_prefix_caching
459
        self.cpu_offload_gb = cpu_offload_gb
460
        self._verify_args()
461
        self._verify_cache_dtype()
462
        self._verify_prefix_caching()
463
464
465
466
467

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

468
    def metrics_info(self):
469
470
        # convert cache_config to dict(key: str, value: str) for prometheus
        # metrics info
471
472
        return {key: str(value) for key, value in self.__dict__.items()}

473
474
475
476
477
478
    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}.")

479
480
481
    def _verify_cache_dtype(self) -> None:
        if self.cache_dtype == "auto":
            pass
482
        elif self.cache_dtype in ("fp8", "fp8_e4m3", "fp8_e5m2"):
483
            logger.info(
484
485
                "Using fp8 data type to store kv cache. It reduces the GPU "
                "memory footprint and boosts the performance. "
486
487
                "Meanwhile, it may cause accuracy drop without a proper "
                "scaling factor")
488
489
490
        else:
            raise ValueError(f"Unknown kv cache dtype: {self.cache_dtype}")

491
492
493
494
495
496
497
498
499
500
501
502
503
    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.")
        if self.cache_dtype == "fp8":
            raise NotImplementedError(
                "Prefix caching is not supported for fp8 cache_dtype. "
                "Run with --kv-cache-dtype auto to use prefix caching.")

504
505
506
507
508
509
510
511
512
513
    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

514
515
516
        msg = (f"{cpu_memory_usage / _GB:.2f} GiB out of "
               f"the {total_cpu_memory / _GB:.2f} GiB total CPU memory is "
               "allocated for the swap space.")
517
518
519
        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:
520
            logger.warning("Possibly too large swap space. %s", msg)
521

522

523
524
525
@dataclass
class TokenizerPoolConfig:
    """Configuration for the tokenizer pool.
526

527
528
529
530
531
532
533
534
    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
535
    pool_type: Union[str, Type["BaseTokenizerGroup"]]
536
537
538
    extra_config: dict

    def __post_init__(self):
539
540
        if self.pool_type not in ("ray", ) and not isinstance(
                self.pool_type, type):
541
542
543
544
545
546
547
548
549
550
            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.
551

552
        If tokenizer_pool_size is 0, return None.
553

554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
        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


576
577
578
579
580
581
582
class LoadFormat(str, enum.Enum):
    AUTO = "auto"
    PT = "pt"
    SAFETENSORS = "safetensors"
    NPCACHE = "npcache"
    DUMMY = "dummy"
    TENSORIZER = "tensorizer"
583
    SHARDED_STATE = "sharded_state"
584
    BITSANDBYTES = "bitsandbytes"
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603


@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.
604
            "bitsandbytes" will load nf4 type weights.
605
606
607
        ignore_patterns: The list of patterns to ignore when loading the model.
            Default to "original/**/*" to avoid repeated loading of llama's 
            checkpoints.
608
            
609
610
611
612
613
614
    """

    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)
615
    ignore_patterns: Optional[Union[List[str], str]] = None
616
617
618
619
620
621
622
623

    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()

624
625
626
627
628
629
630
        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/**/*"]

631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
    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}")


650
class ParallelConfig:
651
652
653
654
655
    """Configuration for the distributed execution.

    Args:
        pipeline_parallel_size: Number of pipeline parallel groups.
        tensor_parallel_size: Number of tensor parallel groups.
656
        worker_use_ray: Deprecated, use distributed_executor_backend instead.
zspo's avatar
zspo committed
657
658
659
        max_parallel_loading_workers: Maximum number of multiple batches
            when load model sequentially. To avoid RAM OOM when using tensor
            parallel and large models.
660
661
        disable_custom_all_reduce: Disable the custom all-reduce kernel and
            fall back to NCCL.
662
663
        tokenizer_pool_config: Config for the tokenizer pool.
            If None, will use synchronous tokenization.
664
665
        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.
666
        placement_group: ray distributed model workers placement group.
667
668
669
670
        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.
671
    """
672

673
674
675
676
    def __init__(
        self,
        pipeline_parallel_size: int,
        tensor_parallel_size: int,
677
        worker_use_ray: Optional[bool] = None,
678
        max_parallel_loading_workers: Optional[int] = None,
679
        disable_custom_all_reduce: bool = False,
680
        tokenizer_pool_config: Optional[TokenizerPoolConfig] = None,
681
        ray_workers_use_nsight: bool = False,
682
        placement_group: Optional["PlacementGroup"] = None,
683
684
        distributed_executor_backend: Optional[Union[
            str, Type["ExecutorBase"]]] = None,
685
686
    ) -> None:
        self.pipeline_parallel_size = pipeline_parallel_size
687
        self.tensor_parallel_size = tensor_parallel_size
688
        self.distributed_executor_backend = distributed_executor_backend
689
        self.max_parallel_loading_workers = max_parallel_loading_workers
690
        self.disable_custom_all_reduce = disable_custom_all_reduce
691
        self.tokenizer_pool_config = tokenizer_pool_config
692
        self.ray_workers_use_nsight = ray_workers_use_nsight
693
        self.placement_group = placement_group
694

695
        self.world_size = pipeline_parallel_size * self.tensor_parallel_size
696
697
698
        if worker_use_ray:
            if self.distributed_executor_backend is None:
                self.distributed_executor_backend = "ray"
699
            elif not self.use_ray:
700
701
702
703
704
                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:
705
706
707
            # We use multiprocessing by default if world_size fits on the
            # current node and we aren't in a ray placement group.

708
            from vllm.executor import ray_utils
709
            backend = "mp"
710
            ray_found = ray_utils.ray_is_available()
711
            if cuda_device_count_stateless() < self.world_size:
712
713
                if not ray_found:
                    raise ValueError("Unable to load Ray which is "
714
715
716
                                     "required for multi-node inference, "
                                     "please install Ray with `pip install "
                                     "ray`.") from ray_utils.ray_import_err
717
718
                backend = "ray"
            elif ray_found:
719
                if self.placement_group:
720
                    backend = "ray"
721
722
723
724
725
726
                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"
727
728
729
            self.distributed_executor_backend = backend
            logger.info("Defaulting to use %s for distributed inference",
                        backend)
730

731
        self._verify_args()
732
        self.rank: int = 0
733

734
735
736
737
738
739
    @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)

740
    def _verify_args(self) -> None:
741
742
743
744
745
746
747
        # 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)):
748
            raise ValueError(
749
750
751
752
                "Unrecognized distributed executor backend "
                f"{self.distributed_executor_backend}. Supported "
                "values are 'ray', 'mp' or custom ExecutorBase subclass.")
        if self.use_ray:
753
754
            from vllm.executor import ray_utils
            ray_utils.assert_ray_available()
755
756
757
758
759
        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.")
760
        if self.ray_workers_use_nsight and not self.use_ray:
761
762
            raise ValueError("Unable to use nsight profiling unless workers "
                             "run with Ray.")
763

764
765

class SchedulerConfig:
766
767
768
769
770
771
772
    """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
773
        max_model_len: Maximum length of a sequence (including prompt
Lily Liu's avatar
Lily Liu committed
774
            and generated text).
775
776
777
778
779
        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.
780
781
        delay_factor: Apply a delay (of delay factor multiplied by previous
            prompt latency) before scheduling next prompt.
782
783
        enable_chunked_prefill: If True, prefill requests can be chunked based
            on the remaining max_num_batched_tokens.
784
        embedding_mode: Whether the running model is for embedding.
785
786
787
788
789
790
        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.
791
    """
792

793
794
795
796
797
798
799
800
801
802
    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:
803
804
805
        if max_num_batched_tokens is not None:
            self.max_num_batched_tokens = max_num_batched_tokens
        else:
806
            if enable_chunked_prefill:
807
808
809
                # 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
810
811
812
813
            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)
814
815
816
817
818
            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:
819
820
            logger.info(
                "Chunked prefill is enabled with max_num_batched_tokens=%d.",
821
                self.max_num_batched_tokens)
822

823
        self.max_num_seqs = max_num_seqs
Lily Liu's avatar
Lily Liu committed
824
        self.max_model_len = max_model_len
825
        self.use_v2_block_manager = use_v2_block_manager
826
827
        self.num_lookahead_slots = num_lookahead_slots
        self.delay_factor = delay_factor
828
        self.chunked_prefill_enabled = enable_chunked_prefill
829
        self.embedding_mode = embedding_mode
830
        self.preemption_mode = preemption_mode
831
832
833
        self._verify_args()

    def _verify_args(self) -> None:
834
835
        if (self.max_num_batched_tokens < self.max_model_len
                and not self.chunked_prefill_enabled):
836
837
838
839
840
841
842
            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.")
843

844
845
846
847
848
        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}).")
849

850
851
852
853
854
855
        if self.num_lookahead_slots < 0:
            raise ValueError(
                "num_lookahead_slots "
                f"({self.num_lookahead_slots}) must be greater than or "
                "equal to 0.")

856

857
class DeviceConfig:
858
    device: Optional[torch.device]
859

860
861
862
    def __init__(self, device: str = "auto") -> None:
        if device == "auto":
            # Automated device type detection
863
            if is_neuron():
864
                self.device_type = "neuron"
865
866
            elif is_openvino():
                self.device_type = "openvino"
867
868
            elif is_tpu():
                self.device_type = "tpu"
869
870
            elif is_cpu():
                self.device_type = "cpu"
871
872
            elif is_xpu():
                self.device_type = "xpu"
873
            else:
874
875
876
                # We don't call torch.cuda.is_available() here to
                # avoid initializing CUDA before workers are forked
                self.device_type = "cuda"
877
878
879
880
881
        else:
            # Device type is assigned explicitly
            self.device_type = device

        # Some device types require processing inputs on CPU
882
        if self.device_type in ["neuron", "openvino"]:
883
            self.device = torch.device("cpu")
884
885
        elif self.device_type in ["tpu"]:
            self.device = None
886
887
888
889
        else:
            # Set device with device type
            self.device = torch.device(self.device_type)

890

891
892
893
894
895
896
897
898
899
900
901
902
903
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],
904
        speculative_draft_tensor_parallel_size: Optional[int],
905
        num_speculative_tokens: Optional[int],
906
907
908
        speculative_max_model_len: Optional[int],
        enable_chunked_prefill: bool,
        use_v2_block_manager: bool,
909
        speculative_disable_by_batch_size: Optional[int],
910
911
        ngram_prompt_lookup_max: Optional[int],
        ngram_prompt_lookup_min: Optional[int],
912
913
914
        draft_token_acceptance_method: str,
        typical_acceptance_sampler_posterior_threshold: Optional[float],
        typical_acceptance_sampler_posterior_alpha: Optional[float],
915
        disable_logprobs: Optional[bool],
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
    ) -> 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.
931
932
            speculative_draft_tensor_parallel_size (Optional[int]): The degree
                of the tensor parallelism for the draft model.
933
            num_speculative_tokens (Optional[int]): The number of speculative
934
935
                tokens, if provided. Will default to the number in the draft
                model config if present, otherwise is required.
936
937
938
939
940
941
942
943
944
            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.
945
946
947
            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.
948
949
950
951
            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.
952
953
954
955
956
957
958
959
960
961
962
963
964
            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.
965
966
967
968
969
            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.
970
    
971
972
973
974
975
        Returns:
            Optional["SpeculativeConfig"]: An instance of SpeculativeConfig if
                the necessary conditions are met, else None.
        """

976
977
978
979
        if speculative_model is None:
            if num_speculative_tokens is not None:
                raise ValueError("num_speculative_tokens was provided without "
                                 "speculative_model.")
980
981
            return None

982
983
984
985
986
987
        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=}")

988
989
990
991
992
993
994
995
996
997
        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.")

998
999
1000
1001
1002
1003
        # 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

1004
1005
        if speculative_model == "[ngram]":
            if ngram_prompt_lookup_min is None:
1006
1007
1008
1009
1010
1011
1012
1013
                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=}")
1014

1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
            # 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,
1036
1037
                max_seq_len_to_capture=target_model_config.
                max_seq_len_to_capture,
1038
1039
1040
                max_logprobs=target_model_config.max_logprobs,
            )

1041
            draft_hf_config = draft_model_config.hf_config
1042

1043
1044
1045
1046
1047
            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)
1048
1049
1050
1051
1052
1053
1054
1055
            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(
1056
1057
1058
                        "This speculative model supports a maximum of "
                        f"num_speculative_tokens={n_predict}, but "
                        f"{num_speculative_tokens=} was provided.")
1059

1060
1061
1062
1063
1064
1065
1066
1067
1068
            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(
1069
1070
                    target_parallel_config,
                    speculative_draft_tensor_parallel_size))
1071

1072
1073
1074
1075
1076
1077
        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.")

1078
1079
1080
1081
        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
1082
1083
        if disable_logprobs is None:
            disable_logprobs = True
1084

1085
1086
1087
1088
        return SpeculativeConfig(
            draft_model_config,
            draft_parallel_config,
            num_speculative_tokens,
1089
            speculative_disable_by_batch_size,
1090
1091
            ngram_prompt_lookup_max,
            ngram_prompt_lookup_min,
1092
1093
1094
1095
1096
            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,
1097
            disable_logprobs=disable_logprobs
1098
1099
        )

1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
    @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,
        )

1135
1136
    @staticmethod
    def create_draft_parallel_config(
1137
1138
1139
        target_parallel_config: ParallelConfig,
        speculative_draft_tensor_parallel_size: Optional[int]
    ) -> ParallelConfig:
1140
1141
        """Create a parallel config for use by the draft worker.

1142
        This is mostly a copy of the target parallel config, except the tp_size.
1143
        """
1144
1145
1146
1147
1148
1149
1150
1151
1152
        if speculative_draft_tensor_parallel_size is None:
            speculative_draft_tensor_parallel_size = \
                  target_parallel_config.tensor_parallel_size
        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")

1153
1154
1155
        draft_parallel_config = ParallelConfig(
            pipeline_parallel_size=target_parallel_config.
            pipeline_parallel_size,
1156
            tensor_parallel_size=speculative_draft_tensor_parallel_size,
1157
1158
            distributed_executor_backend=target_parallel_config.
            distributed_executor_backend,
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
            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,
1176
1177
1178
        speculative_disable_by_batch_size: Optional[int],
        ngram_prompt_lookup_max: Optional[int],
        ngram_prompt_lookup_min: Optional[int],
1179
1180
1181
        draft_token_acceptance_method: str,
        typical_acceptance_sampler_posterior_threshold: float,
        typical_acceptance_sampler_posterior_alpha: float,
1182
        disable_logprobs: bool,
1183
1184
1185
1186
1187
1188
1189
1190
    ):
        """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.
1191
1192
1193
1194
1195
            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.
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
            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.
1209
1210
1211
1212
1213
1214
            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.
1215
1216
1217
1218
        """
        self.draft_model_config = draft_model_config
        self.draft_parallel_config = draft_parallel_config
        self.num_speculative_tokens = num_speculative_tokens
1219
1220
1221
1222
        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
1223
1224
1225
1226
1227
        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
1228
        self.disable_logprobs = disable_logprobs
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239

        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)
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
            # 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}")
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276

    @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:
1277
1278
1279
1280
        if self.ngram_prompt_lookup_max > 0:
            draft_model = "[ngram]"
        else:
            draft_model = self.draft_model_config.model
1281
1282
1283
1284
        num_spec_tokens = self.num_speculative_tokens
        return f"SpeculativeConfig({draft_model=}, {num_spec_tokens=})"


1285
1286
1287
1288
@dataclass
class LoRAConfig:
    max_lora_rank: int
    max_loras: int
1289
    fully_sharded_loras: bool = False
1290
1291
1292
1293
1294
    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
1295
    long_lora_scaling_factors: Optional[Tuple[float]] = None
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315

    def __post_init__(self):
        # Keep this in sync with csrc/punica/bgmv/bgmv_config.h
        possible_max_ranks = (8, 16, 32, 64)
        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
1316
                f"max_loras ({self.max_loras})")
1317
1318
1319
1320
1321
1322

    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)
1323
1324
1325
1326
        if model_config.quantization and model_config.quantization not in [
                "awq", "gptq"
        ]:
            # TODO support marlin and squeezellm
1327
1328
            logger.warning("%s quantization is not tested with LoRA yet.",
                           model_config.quantization)
1329
1330
1331
1332
1333
1334
1335

    def verify_with_scheduler_config(self, scheduler_config: SchedulerConfig):
        if scheduler_config.max_num_batched_tokens > 65528:
            raise ValueError(
                "Due to limitations of the custom LoRA CUDA kernel, "
                "max_num_batched_tokens must be <= 65528 when "
                "LoRA is enabled.")
1336
1337
        if scheduler_config.chunked_prefill_enabled:
            raise ValueError("LoRA is not supported with chunked prefill yet.")
1338
1339


1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
@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)


1373
@dataclass
1374
class MultiModalConfig:
1375
    """Configs the input data format and how models should run for
1376
1377
1378
    multimodal models."""
    # TODO: Add configs to init vision tower or not.
    pass
1379

1380

1381
1382
1383
1384
1385
1386
1387
1388
_STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.float16,
    "float16": torch.float16,
    "float": torch.float32,
    "float32": torch.float32,
    "bfloat16": torch.bfloat16,
}

1389
_ROCM_NOT_SUPPORTED_DTYPE: List[str] = []  #
1390

1391
1392
1393

def _get_and_verify_dtype(
    config: PretrainedConfig,
1394
    dtype: Union[str, torch.dtype],
1395
1396
1397
1398
1399
1400
1401
) -> 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

1402
1403
1404
1405
    if isinstance(dtype, str):
        dtype = dtype.lower()
        if dtype == "auto":
            if config_dtype == torch.float32:
Woosuk Kwon's avatar
Woosuk Kwon committed
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
                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
1416
1417
            else:
                torch_dtype = config_dtype
1418
        else:
1419
1420
1421
1422
1423
            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
1424
    else:
1425
        raise ValueError(f"Unknown dtype: {dtype}")
1426
1427
1428
1429
1430

    # Verify the dtype.
    if torch_dtype != config_dtype:
        if torch_dtype == torch.float32:
            # Upcasting to float32 is allowed.
1431
            logger.info("Upcasting %s to %s.", config_dtype, torch_dtype)
1432
1433
1434
            pass
        elif config_dtype == torch.float32:
            # Downcasting from float32 to float16 or bfloat16 is allowed.
1435
            logger.info("Downcasting %s to %s.", config_dtype, torch_dtype)
1436
1437
            pass
        else:
Woosuk Kwon's avatar
Woosuk Kwon committed
1438
            # Casting between float16 and bfloat16 is allowed with a warning.
1439
            logger.warning("Casting %s to %s.", config_dtype, torch_dtype)
1440
1441

    return torch_dtype
1442
1443
1444
1445
1446


def _get_and_verify_max_len(
    hf_config: PretrainedConfig,
    max_model_len: Optional[int],
1447
1448
    disable_sliding_window: bool,
    sliding_window_len: Optional[int],
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
) -> 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",
1459
1460
        # ChatGLM2
        "seq_length",
1461
1462
        # Command-R
        "model_max_length",
1463
1464
1465
1466
1467
        # Others
        "max_sequence_length",
        "max_seq_length",
        "seq_len",
    ]
1468
    # Choose the smallest "max_length" from the possible keys.
1469
    max_len_key = None
1470
    for key in possible_keys:
1471
1472
1473
1474
1475
        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)
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485

    # 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.
1486
    if derived_max_model_len == float("inf"):
1487
1488
1489
1490
1491
1492
1493
1494
        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: "
1495
            "%s. Assuming the model's maximum length is %d.", possible_keys,
1496
            default_max_len)
1497
        derived_max_model_len = default_max_len
1498

1499
    rope_scaling = getattr(hf_config, "rope_scaling", None)
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
    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
1526

1527
1528
    # If the user specified a max length, make sure it is smaller than the
    # derived length from the HF model config.
1529
    if max_model_len is None:
1530
        max_model_len = int(derived_max_model_len)
1531
    elif max_model_len > derived_max_model_len:
1532
1533
1534
1535
1536
        # 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:
1537
1538
1539
1540
1541
1542
1543
            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.")
1544
1545
1546
1547
1548
1549
1550
1551
1552
            pass
        else:
            raise ValueError(
                f"User-specified max_model_len ({max_model_len}) is greater "
                "than the derived max_model_len "
                f"({max_len_key}={derived_max_model_len} or model_max_length="
                f"{model_max_length} in model's config.json). This may lead "
                "to incorrect model outputs or CUDA errors. Make sure the "
                "value is correct and within the model context size.")
1553
    return int(max_model_len)
1554
1555


1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
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


1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
@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}")


1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
@dataclass
class ObservabilityConfig:
    """Configuration for observability."""
    otlp_traces_endpoint: Optional[str] = None

    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'")


1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
@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
1609
    load_config: LoadConfig
1610
    lora_config: Optional[LoRAConfig]
1611
    multimodal_config: Optional[MultiModalConfig]
1612
    speculative_config: Optional[SpeculativeConfig]
1613
    decoding_config: Optional[DecodingConfig]
1614
    observability_config: Optional[ObservabilityConfig]
1615
    prompt_adapter_config: Optional[PromptAdapterConfig]
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626

    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)
1627
1628
1629
        if self.prompt_adapter_config:
            self.prompt_adapter_config.verify_with_model_config(
                self.model_config)
1630
1631
1632
1633
1634
1635

    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))