config.py 29.5 KB
Newer Older
1
from typing import TYPE_CHECKING, Optional, Union, ClassVar
2
from dataclasses import dataclass
3
import os
4
from packaging.version import Version
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.transformers_utils.config import get_config
11
from vllm.utils import get_cpu_memory, is_hip, is_neuron, get_nvcc_cuda_version
12

13
14
15
if TYPE_CHECKING:
    from ray.util.placement_group import PlacementGroup

16
17
logger = init_logger(__name__)

18
_GB = 1 << 30
19

20
21

class ModelConfig:
22
23
24
25
    """Configuration for the model.

    Args:
        model: Name or path of the huggingface model to use.
26
        tokenizer: Name or path of the huggingface tokenizer to use.
27
28
        tokenizer_mode: Tokenizer mode. "auto" will use the fast tokenizer if
            available, and "slow" will always use the slow tokenizer.
29
30
        trust_remote_code: Trust remote code (e.g., from HuggingFace) when
            downloading the model and tokenizer.
31
32
        download_dir: Directory to download and load the weights, default to the
            default cache directory of huggingface.
33
34
35
36
37
38
39
40
41
42
        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.
43
44
45
46
        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
47
48
49
        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.
50
        code_revision: The specific revision to use for the model code on
51
            Hugging Face Hub. It can be a branch name, a tag name, or a
52
            commit id. If unspecified, will use the default version.
53
54
55
        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.
56
57
        max_model_len: Maximum length of a sequence (including prompt and
            output). If None, will be derived from the model.
58
59
        quantization: Quantization method that was used to quantize the model
            weights. If None, we assume the model weights are not quantized.
60
61
62
63
64
65
        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
            to eager mode.
66
    """
67
68
69
70

    def __init__(
        self,
        model: str,
71
72
        tokenizer: str,
        tokenizer_mode: str,
73
        trust_remote_code: bool,
74
        download_dir: Optional[str],
75
        load_format: str,
76
        dtype: Union[str, torch.dtype],
77
        seed: int,
78
        revision: Optional[str] = None,
79
        code_revision: Optional[str] = None,
80
        tokenizer_revision: Optional[str] = None,
81
        max_model_len: Optional[int] = None,
82
        quantization: Optional[str] = None,
83
84
        enforce_eager: bool = False,
        max_context_len_to_capture: Optional[int] = None,
85
        max_logprobs: int = 5,
86
87
    ) -> None:
        self.model = model
88
        self.tokenizer = tokenizer
89
        self.tokenizer_mode = tokenizer_mode
90
        self.trust_remote_code = trust_remote_code
91
        self.download_dir = download_dir
92
        self.load_format = load_format
93
        self.seed = seed
Jasmond L's avatar
Jasmond L committed
94
        self.revision = revision
95
        self.code_revision = code_revision
96
        self.tokenizer_revision = tokenizer_revision
97
        self.quantization = quantization
98
99
        self.enforce_eager = enforce_eager
        self.max_context_len_to_capture = max_context_len_to_capture
100
        self.max_logprobs = max_logprobs
101

102
103
104
105
        if os.environ.get("VLLM_USE_MODELSCOPE", "False").lower() == "true":
            # download model from ModelScope hub,
            # lazy import so that modelscope is not required for normal use.
            from modelscope.hub.snapshot_download import snapshot_download  # pylint: disable=C
106

107
108
109
110
111
112
            if not os.path.exists(model):
                model_path = snapshot_download(model_id=model,
                                               cache_dir=download_dir,
                                               revision=revision)
            else:
                model_path = model
113
114
115
116
            self.model = model_path
            self.download_dir = model_path
            self.tokenizer = model_path

117
118
        self.hf_config = get_config(self.model, trust_remote_code, revision,
                                    code_revision)
119
        self.dtype = _get_and_verify_dtype(self.hf_config, dtype)
120
121
        self.max_model_len = _get_and_verify_max_len(self.hf_config,
                                                     max_model_len)
122
        self._verify_load_format()
123
        self._verify_tokenizer_mode()
124
        self._verify_quantization()
125
        self._verify_cuda_graph()
126

127
128
    def _verify_load_format(self) -> None:
        load_format = self.load_format.lower()
129
130
131
        supported_load_format = [
            "auto", "pt", "safetensors", "npcache", "dummy"
        ]
kliuae's avatar
kliuae committed
132
        rocm_not_supported_load_format = []
133
        if load_format not in supported_load_format:
134
135
136
            raise ValueError(
                f"Unknown load format: {self.load_format}. Must be one of "
                "'auto', 'pt', 'safetensors', 'npcache', or 'dummy'.")
kliuae's avatar
kliuae committed
137
138
139
140
141
142
        if is_hip() and load_format in rocm_not_supported_load_format:
            rocm_supported_load_format = [
                f for f in supported_load_format
                if (f not in rocm_not_supported_load_format)
            ]
            raise ValueError(
143
                f"load format '{load_format}' is not supported in ROCm. "
kliuae's avatar
kliuae committed
144
145
                f"Supported load format are "
                f"{rocm_supported_load_format}")
146

147
        # TODO: Remove this check once HF updates the pt weights of Mixtral.
148
        architectures = getattr(self.hf_config, "architectures", [])
Roy's avatar
Roy committed
149
150
151
152
        if "MixtralForCausalLM" in architectures and load_format == "pt":
            raise ValueError(
                "Currently, the 'pt' format is not supported for Mixtral. "
                "Please use the 'safetensors' format instead. ")
153
154
        self.load_format = load_format

155
156
157
158
159
160
161
    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
162

163
    def _verify_quantization(self) -> None:
164
165
        supported_quantization = ["awq", "gptq", "squeezellm", "marlin"]
        rocm_not_supported_quantization = ["awq", "marlin"]
166
167
168
169
170
171
172
        if self.quantization is not None:
            self.quantization = self.quantization.lower()

        # Parse quantization method from the HF model config, if available.
        hf_quant_config = getattr(self.hf_config, "quantization_config", None)
        if hf_quant_config is not None:
            hf_quant_method = str(hf_quant_config["quant_method"]).lower()
173

174
175
176
177
            # If the GPTQ model is serialized in marlin format, use marlin.
            if (hf_quant_method == "gptq"
                    and "is_marlin_format" in hf_quant_config
                    and hf_quant_config["is_marlin_format"]):
178
179
                logger.info("The model is serialized in Marlin format. "
                            "Using Marlin kernel.")
180
                hf_quant_method = "marlin"
181
182
183
                if self.quantization == "gptq":
                    self.quantization = hf_quant_method

184
185
186
187
188
189
190
191
192
193
194
195
196
197
            if self.quantization is None:
                self.quantization = hf_quant_method
            elif self.quantization != hf_quant_method:
                raise ValueError(
                    "Quantization method specified in the model config "
                    f"({hf_quant_method}) does not match the quantization "
                    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}.")
198
199
200
            if is_hip(
            ) and self.quantization in rocm_not_supported_quantization:
                raise ValueError(
201
202
                    f"{self.quantization} quantization is currently not "
                    f"supported in ROCm.")
203
204
205
206
207
            if self.quantization != "marlin":
                logger.warning(
                    f"{self.quantization} quantization is not fully "
                    "optimized yet. The speed can be slower than "
                    "non-quantized models.")
208

209
210
211
212
213
214
    def _verify_cuda_graph(self) -> None:
        if self.max_context_len_to_capture is None:
            self.max_context_len_to_capture = self.max_model_len
        self.max_context_len_to_capture = min(self.max_context_len_to_capture,
                                              self.max_model_len)

215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
        total_num_attention_heads = self.hf_config.num_attention_heads
        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}).")

        total_num_hidden_layers = self.hf_config.num_hidden_layers
        pipeline_parallel_size = parallel_config.pipeline_parallel_size
        if total_num_hidden_layers % pipeline_parallel_size != 0:
            raise ValueError(
                f"Total number of hidden layers ({total_num_hidden_layers}) "
                "must be divisible by pipeline parallel size "
                f"({pipeline_parallel_size}).")

235
    def get_sliding_window(self) -> Optional[int]:
236
237
238
239
240
241
242
243
244
        """Get the sliding window size, or None if disabled.
        """

        # 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.
        if (hasattr(self.hf_config, "use_sliding_window")
                and not self.hf_config.use_sliding_window):
            return None
245
246
247
248
249
        return getattr(self.hf_config, "sliding_window", None)

    def get_vocab_size(self) -> int:
        return self.hf_config.vocab_size

250
251
252
253
    def get_hidden_size(self) -> int:
        return self.hf_config.hidden_size

    def get_head_size(self) -> int:
254
255
        if hasattr(self.hf_config, "head_dim"):
            return self.hf_config.head_dim
256
257
258
        # FIXME(woosuk): This may not be true for all models.
        return self.hf_config.hidden_size // self.hf_config.num_attention_heads

259
260
    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
Zhuohan Li's avatar
Zhuohan Li committed
261
        # For GPTBigCode & Falcon:
262
        # NOTE: for falcon, when new_decoder_architecture is True, the
Zhuohan Li's avatar
Zhuohan Li committed
263
264
        # multi_query flag is ignored and we use n_head_kv for the number of
        # KV heads.
265
        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
266
        new_decoder_arch_falcon = (
267
            self.hf_config.model_type in falcon_model_types
268
269
270
            and getattr(self.hf_config, "new_decoder_architecture", False))
        if not new_decoder_arch_falcon and getattr(self.hf_config,
                                                   "multi_query", False):
Zhuohan Li's avatar
Zhuohan Li committed
271
            # Multi-query attention, only one KV head.
Woosuk Kwon's avatar
Woosuk Kwon committed
272
            # Currently, tensor parallelism is not supported in this case.
Zhuohan Li's avatar
Zhuohan Li committed
273
            return 1
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301

        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:
            num_kv_heads = getattr(self.hf_config, attr, None)
            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.
        return self.hf_config.num_attention_heads

    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)
302
303
304
305
306
307
308

    def get_num_layers(self, parallel_config: "ParallelConfig") -> int:
        total_num_hidden_layers = self.hf_config.num_hidden_layers
        return total_num_hidden_layers // parallel_config.pipeline_parallel_size


class CacheConfig:
309
310
311
312
313
    """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
314
            vLLM execution.
315
        swap_space: Size of the CPU swap space per GPU (in GiB).
316
        cache_dtype: Data type for kv cache storage.
317
    """
318

319
320
321
322
323
    def __init__(
        self,
        block_size: int,
        gpu_memory_utilization: float,
        swap_space: int,
324
        cache_dtype: str,
325
        sliding_window: Optional[int] = None,
326
        enable_prefix_caching: bool = False,
327
328
329
    ) -> None:
        self.block_size = block_size
        self.gpu_memory_utilization = gpu_memory_utilization
330
        self.swap_space_bytes = swap_space * _GB
331
        self.cache_dtype = cache_dtype
332
        self.sliding_window = sliding_window
333
        self.enable_prefix_caching = enable_prefix_caching
334
        self._verify_args()
335
        self._verify_cache_dtype()
336
337
338
339
340

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

341
    def metrics_info(self):
342
343
        # convert cache_config to dict(key: str, value: str) for prometheus
        # metrics info
344
345
        return {key: str(value) for key, value in self.__dict__.items()}

346
347
348
349
350
351
    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}.")

352
353
354
355
    def _verify_cache_dtype(self) -> None:
        if self.cache_dtype == "auto":
            pass
        elif self.cache_dtype == "fp8_e5m2":
356
357
358
            if is_hip():
                raise NotImplementedError(
                    "FP8_E5M2 KV Cache on AMD GPU has not been supported yet.")
359
            nvcc_cuda_version = get_nvcc_cuda_version()
360
            if nvcc_cuda_version and nvcc_cuda_version < Version("11.8"):
361
362
363
364
365
366
367
368
369
370
371
372
                raise ValueError(
                    "FP8 is not supported when cuda version is lower than 11.8."
                )
            logger.info(
                "Using fp8_e5m2 data type to store kv cache. It reduces "
                "the GPU memory footprint and boosts the performance. "
                "But it may cause slight accuracy drop. "
                "Currently we only support fp8 without scaling factors and "
                "make e5m2 as a default format.")
        else:
            raise ValueError(f"Unknown kv cache dtype: {self.cache_dtype}")

373
374
375
376
377
378
379
380
381
382
    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

383
384
385
        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.")
386
387
388
        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:
389
            logger.warning("Possibly too large swap space. " + msg)
390

391
392

class ParallelConfig:
393
394
395
396
397
398
399
400
    """Configuration for the distributed execution.

    Args:
        pipeline_parallel_size: Number of pipeline parallel groups.
        tensor_parallel_size: Number of tensor parallel groups.
        worker_use_ray: Whether to use Ray for model workers. Will be set to
            True if either pipeline_parallel_size or tensor_parallel_size is
            greater than 1.
zspo's avatar
zspo committed
401
402
403
        max_parallel_loading_workers: Maximum number of multiple batches
            when load model sequentially. To avoid RAM OOM when using tensor
            parallel and large models.
404
405
        disable_custom_all_reduce: Disable the custom all-reduce kernel and
            fall back to NCCL.
406
407
        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.
408
    """
409

410
411
412
413
    def __init__(
        self,
        pipeline_parallel_size: int,
        tensor_parallel_size: int,
414
        worker_use_ray: bool,
415
        max_parallel_loading_workers: Optional[int] = None,
416
        disable_custom_all_reduce: bool = False,
417
        ray_workers_use_nsight: bool = False,
418
        placement_group: Optional["PlacementGroup"] = None,
419
420
    ) -> None:
        self.pipeline_parallel_size = pipeline_parallel_size
421
        if is_neuron():
422
423
424
            # For Neuron device support, here we assign TP=1 to avoid sharding
            # within vLLM directly. Transformer-neuronx would take
            # neuron_tp_degree attribute, and distribute the workload
425
426
427
428
429
            # to multiple NeuronCores.
            self.tensor_parallel_size = 1
            self.neuron_tp_degree = tensor_parallel_size
        else:
            self.tensor_parallel_size = tensor_parallel_size
430
        self.worker_use_ray = worker_use_ray
431
        self.max_parallel_loading_workers = max_parallel_loading_workers
432
        self.disable_custom_all_reduce = disable_custom_all_reduce
433
        self.ray_workers_use_nsight = ray_workers_use_nsight
434
        self.placement_group = placement_group
435

436
437
438
        self.world_size = pipeline_parallel_size * self.tensor_parallel_size
        # Ray worker is not supported for Neuron backend.
        if self.world_size > 1 and not is_neuron():
439
            self.worker_use_ray = True
440
441
442
443
444
445
        self._verify_args()

    def _verify_args(self) -> None:
        if self.pipeline_parallel_size > 1:
            raise NotImplementedError(
                "Pipeline parallelism is not supported yet.")
446
447
448
449
450
451
452
453
454
455
456
        if not self.disable_custom_all_reduce and self.world_size > 1:
            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.")
            elif self.pipeline_parallel_size > 1:
                self.disable_custom_all_reduce = True
                logger.info(
                    "Disabled the custom all-reduce kernel because it is not "
                    "supported with pipeline parallelism.")
457
458
459
        if self.ray_workers_use_nsight and not self.worker_use_ray:
            raise ValueError("Unable to use nsight profiling unless workers "
                             "run with Ray.")
460
461
462
463

        # FIXME(woosuk): Fix the stability issues and re-enable the custom
        # all-reduce kernel.
        if not self.disable_custom_all_reduce and self.world_size > 1:
464
465
            self.disable_custom_all_reduce = True
            logger.info(
466
467
468
                "Custom all-reduce kernels are temporarily disabled due to "
                "stability issues. We will re-enable them once the issues are "
                "resolved.")
469
470
471


class SchedulerConfig:
472
473
474
475
476
477
478
    """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
479
        max_model_len: Maximum length of a sequence (including prompt
Lily Liu's avatar
Lily Liu committed
480
            and generated text).
481
        max_paddings: Maximum number of paddings to be added to a batch.
482
    """
483

484
485
486
487
488
    def __init__(
        self,
        max_num_batched_tokens: Optional[int],
        max_num_seqs: int,
        max_model_len: int,
489
        max_paddings: int,
490
491
492
493
494
495
496
    ) -> None:
        if max_num_batched_tokens is not None:
            self.max_num_batched_tokens = max_num_batched_tokens
        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)
497
        self.max_num_seqs = max_num_seqs
Lily Liu's avatar
Lily Liu committed
498
        self.max_model_len = max_model_len
499
        self.max_paddings = max_paddings
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
        self._verify_args()

    def _verify_args(self) -> None:
        if self.max_num_batched_tokens < self.max_model_len:
            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.")
        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}).")
516
517


518
519
class DeviceConfig:

520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
    def __init__(self, device: str = "auto") -> None:
        if device == "auto":
            # Automated device type detection
            if torch.cuda.is_available():
                self.device_type = "cuda"
            elif is_neuron():
                self.device_type = "neuron"
            else:
                raise RuntimeError("No supported device detected.")
        else:
            # Device type is assigned explicitly
            self.device_type = device

        # Some device types require processing inputs on CPU
        if self.device_type in ["neuron"]:
            self.device = torch.device("cpu")
        else:
            # Set device with device type
            self.device = torch.device(self.device_type)

    @property
    def is_neuron(self):
        return self.device_type == "neuron"
543
544


545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
@dataclass
class LoRAConfig:
    max_lora_rank: int
    max_loras: int
    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

    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
574
                f"max_loras ({self.max_loras})")
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592

    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)
        if model_config.quantization is not None:
            raise ValueError(
                "LoRA is not supported with quantized models yet.")

    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.")


593
594
595
596
597
598
599
600
_STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.float16,
    "float16": torch.float16,
    "float": torch.float32,
    "float32": torch.float32,
    "bfloat16": torch.bfloat16,
}

601
602
_ROCM_NOT_SUPPORTED_DTYPE = ["float", "float32"]

603
604
605

def _get_and_verify_dtype(
    config: PretrainedConfig,
606
    dtype: Union[str, torch.dtype],
607
608
609
610
611
612
613
) -> 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

614
615
616
617
618
619
620
621
622
    if isinstance(dtype, str):
        dtype = dtype.lower()
        if dtype == "auto":
            if config_dtype == torch.float32:
                # Following the common practice, we use float16 for float32
                # models.
                torch_dtype = torch.float16
            else:
                torch_dtype = config_dtype
623
        else:
624
625
626
627
628
            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
629
    else:
630
        raise ValueError(f"Unknown dtype: {dtype}")
631

632
633
634
635
636
    if is_hip() and torch_dtype == torch.float32:
        rocm_supported_dtypes = [
            k for k, v in _STR_DTYPE_TO_TORCH_DTYPE.items()
            if (k not in _ROCM_NOT_SUPPORTED_DTYPE)
        ]
637
        raise ValueError(f"dtype '{dtype}' is not supported in ROCm. "
638
639
                         f"Supported dtypes are {rocm_supported_dtypes}")

640
641
642
643
644
645
646
647
648
    # Verify the dtype.
    if torch_dtype != config_dtype:
        if torch_dtype == torch.float32:
            # Upcasting to float32 is allowed.
            pass
        elif config_dtype == torch.float32:
            # Downcasting from float32 to float16 or bfloat16 is allowed.
            pass
        else:
Woosuk Kwon's avatar
Woosuk Kwon committed
649
            # Casting between float16 and bfloat16 is allowed with a warning.
650
            logger.warning(f"Casting {config_dtype} to {torch_dtype}.")
651
652

    return torch_dtype
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667


def _get_and_verify_max_len(
    hf_config: PretrainedConfig,
    max_model_len: Optional[int],
) -> 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",
668
669
        # ChatGLM2
        "seq_length",
670
671
672
673
674
675
676
677
678
        # Others
        "max_sequence_length",
        "max_seq_length",
        "seq_len",
    ]
    for key in possible_keys:
        max_len_key = getattr(hf_config, key, None)
        if max_len_key is not None:
            derived_max_model_len = min(derived_max_model_len, max_len_key)
679
    if derived_max_model_len == float("inf"):
680
681
682
683
684
685
686
687
688
689
690
        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: "
            f"{possible_keys}. Assuming the model's maximum length is "
            f"{default_max_len}.")
        derived_max_model_len = default_max_len
691

692
693
694
695
    rope_scaling = getattr(hf_config, "rope_scaling", None)
    if rope_scaling is not None:
        assert "factor" in rope_scaling
        scaling_factor = rope_scaling["factor"]
Antoni Baum's avatar
Antoni Baum committed
696
697
698
        if rope_scaling["type"] == "yarn":
            derived_max_model_len = rope_scaling[
                "original_max_position_embeddings"]
699
700
        derived_max_model_len *= scaling_factor

701
702
703
704
705
706
707
708
709
    if max_model_len is None:
        max_model_len = derived_max_model_len
    elif max_model_len > derived_max_model_len:
        raise ValueError(
            f"User-specified max_model_len ({max_model_len}) is greater than "
            f"the derived max_model_len ({max_len_key}={derived_max_model_len}"
            " 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.")
710
    return int(max_model_len)