config.py

import enum
import json
import os
from dataclasses import dataclass
from typing import TYPE_CHECKING, ClassVar, Optional, Union

import torch
from packaging.version import Version
from transformers import PretrainedConfig

from vllm.logger import init_logger
from vllm.transformers_utils.config import get_config, get_hf_text_config
from vllm.utils import (get_cpu_memory, get_nvcc_cuda_version, is_cpu, is_hip,
                        is_neuron)

if TYPE_CHECKING:
    from ray.util.placement_group import PlacementGroup

logger = init_logger(__name__)

_GB = 1 << 30


class ModelConfig:
    """Configuration for the model.

    Args:
        model: Name or path of the huggingface model to use.
        tokenizer: Name or path of the huggingface tokenizer to use.
        tokenizer_mode: Tokenizer mode. "auto" will use the fast tokenizer if
            available, and "slow" will always use the slow tokenizer.
        trust_remote_code: Trust remote code (e.g., from HuggingFace) when
            downloading the model and tokenizer.
        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.
        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.
        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.
        code_revision: The specific revision to use for the model code on
            Hugging Face Hub. It can be a branch name, a tag name, or a
            commit id. If unspecified, will use the default version.
        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.
        max_model_len: Maximum length of a sequence (including prompt and
            output). If None, will be derived from the model.
        quantization: Quantization method that was used to quantize the model
            weights. If None, we assume the model weights are not quantized.
        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.
    """

    def __init__(
        self,
        model: str,
        tokenizer: str,
        tokenizer_mode: str,
        trust_remote_code: bool,
        download_dir: Optional[str],
        load_format: str,
        dtype: Union[str, torch.dtype],
        seed: int,
        revision: Optional[str] = None,
        code_revision: Optional[str] = None,
        tokenizer_revision: Optional[str] = None,
        max_model_len: Optional[int] = None,
        quantization: Optional[str] = None,
        enforce_eager: bool = False,
        max_context_len_to_capture: Optional[int] = None,
        max_logprobs: int = 5,
    ) -> None:
        self.model = model
        self.tokenizer = tokenizer
        self.tokenizer_mode = tokenizer_mode
        self.trust_remote_code = trust_remote_code
        self.download_dir = download_dir
        self.load_format = load_format
        self.seed = seed
        self.revision = revision
        self.code_revision = code_revision
        self.tokenizer_revision = tokenizer_revision
        self.quantization = quantization
        self.enforce_eager = enforce_eager
        self.max_context_len_to_capture = max_context_len_to_capture
        self.max_logprobs = max_logprobs

        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.
            # pylint: disable=C.
            from modelscope.hub.snapshot_download import snapshot_download

            if not os.path.exists(model):
                model_path = snapshot_download(model_id=model,
                                               cache_dir=download_dir,
                                               revision=revision)
            else:
                model_path = model
            self.model = model_path
            self.download_dir = model_path
            self.tokenizer = model_path

        self.hf_config = get_config(self.model, trust_remote_code, revision,
                                    code_revision)
        self.hf_text_config = get_hf_text_config(self.hf_config)
        self.dtype = _get_and_verify_dtype(self.hf_text_config, dtype)
        self.max_model_len = _get_and_verify_max_len(self.hf_text_config,
                                                     max_model_len)
        self._verify_load_format()
        self._verify_tokenizer_mode()
        self._verify_quantization()
        self._verify_cuda_graph()

    def _verify_load_format(self) -> None:
        load_format = self.load_format.lower()
        supported_load_format = [
            "auto", "pt", "safetensors", "npcache", "dummy"
        ]
        rocm_not_supported_load_format = []
        if load_format not in supported_load_format:
            raise ValueError(
                f"Unknown load format: {self.load_format}. Must be one of "
                "'auto', 'pt', 'safetensors', 'npcache', or 'dummy'.")
        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(
                f"load format '{load_format}' is not supported in ROCm. "
                f"Supported load format are "
                f"{rocm_supported_load_format}")

        # TODO: Remove this check once HF updates the pt weights of Mixtral.
        architectures = getattr(self.hf_config, "architectures", [])
        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. ")
        self.load_format = load_format

    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

    def _verify_quantization(self) -> None:
        supported_quantization = ["awq", "gptq", "squeezellm", "marlin"]
        rocm_not_supported_quantization = ["awq", "marlin"]
        if self.quantization is not None:
            self.quantization = self.quantization.lower()

        # Parse quantization method from the HF model config, if available.
        quant_cfg = getattr(self.hf_config, "quantization_config", None)
        if quant_cfg is not None:
            quant_method = quant_cfg.get("quant_method", "").lower()
            # compat: autogptq >=0.8.0 use checkpoint_format: str
            # compat: autogptq <=0.7.1 is_marlin_format: bool
            is_format_marlin = (quant_cfg.get("checkpoint_format") == "marlin"
                                or quant_cfg.get("is_marlin_format", False))

            # Use marlin if the GPTQ model is serialized in marlin format.
            if quant_method == "gptq" and is_format_marlin:
                logger.info("The model is serialized in Marlin format. "
                            "Using Marlin kernel.")
                quant_method = "marlin"
                if self.quantization == "gptq":
                    self.quantization = quant_method

            if self.quantization is None:
                self.quantization = quant_method
            elif self.quantization != quant_method:
                raise ValueError(
                    "Quantization method specified in the model config "
                    f"({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}.")
            if is_hip(
            ) and self.quantization in rocm_not_supported_quantization:
                raise ValueError(
                    f"{self.quantization} quantization is currently not "
                    f"supported in ROCm.")
            if self.quantization != "marlin":
                logger.warning(
                    f"{self.quantization} quantization is not fully "
                    "optimized yet. The speed can be slower than "
                    "non-quantized models.")

    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)

    def verify_with_parallel_config(
        self,
        parallel_config: "ParallelConfig",
    ) -> None:
        total_num_attention_heads = self.hf_text_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_text_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}).")

    def get_sliding_window(self) -> Optional[int]:
        """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_text_config, "use_sliding_window")
                and not self.hf_text_config.use_sliding_window):
            return None
        return getattr(self.hf_text_config, "sliding_window", None)

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

    def get_hidden_size(self) -> int:
        return self.hf_text_config.hidden_size

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

    def get_total_num_kv_heads(self) -> int:
        """Returns the total number of KV heads."""
        # For GPTBigCode & Falcon:
        # NOTE: for falcon, when new_decoder_architecture is True, the
        # multi_query flag is ignored and we use n_head_kv for the number of
        # KV heads.
        falcon_model_types = ["falcon", "RefinedWeb", "RefinedWebModel"]
        new_decoder_arch_falcon = (
            self.hf_config.model_type in falcon_model_types
            and getattr(self.hf_config, "new_decoder_architecture", False))
        if not new_decoder_arch_falcon and getattr(self.hf_text_config,
                                                   "multi_query", False):
            # Multi-query attention, only one KV head.
            # Currently, tensor parallelism is not supported in this case.
            return 1

        # For DBRX and MPT
        if self.hf_config.model_type in ["dbrx", "mpt"]:
            return getattr(self.hf_config.attn_config, "kv_n_heads",
                           self.hf_config.num_attention_heads)

        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_text_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_text_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)

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


class CacheConfig:
    """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
            vLLM execution.
        swap_space: Size of the CPU swap space per GPU (in GiB).
        cache_dtype: Data type for kv cache storage.
        forced_num_gpu_blocks: Number of GPU blocks to use. This overrides the
            profiled num_gpu_blocks if specified. Does nothing if None.
    """

    def __init__(
        self,
        block_size: int,
        gpu_memory_utilization: float,
        swap_space: int,
        cache_dtype: str,
        forced_num_gpu_blocks: Optional[int] = None,
        sliding_window: Optional[int] = None,
        enable_prefix_caching: bool = False,
    ) -> None:
        self.block_size = block_size
        self.gpu_memory_utilization = gpu_memory_utilization
        self.swap_space_bytes = swap_space * _GB
        self.forced_num_gpu_blocks = forced_num_gpu_blocks
        self.cache_dtype = cache_dtype
        self.sliding_window = sliding_window
        self.enable_prefix_caching = enable_prefix_caching
        self._verify_args()
        self._verify_cache_dtype()

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

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

    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}.")

    def _verify_cache_dtype(self) -> None:
        if self.cache_dtype == "auto":
            pass
        elif self.cache_dtype == "fp8_e5m2":
            if is_hip():
                raise NotImplementedError(
                    "FP8_E5M2 KV Cache on AMD GPU has not been supported yet.")
            nvcc_cuda_version = get_nvcc_cuda_version()
            if nvcc_cuda_version and nvcc_cuda_version < Version("11.8"):
                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}")

    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

        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.")
        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:
            logger.warning("Possibly too large swap space. " + msg)


@dataclass
class TokenizerPoolConfig:
    """Configuration for the tokenizer pool.
    
    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
    pool_type: str
    extra_config: dict

    def __post_init__(self):
        if self.pool_type not in ("ray", ):
            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.
        
        If tokenizer_pool_size is 0, return None.
        
        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


class ParallelConfig:
    """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.
        max_parallel_loading_workers: Maximum number of multiple batches
            when load model sequentially. To avoid RAM OOM when using tensor
            parallel and large models.
        disable_custom_all_reduce: Disable the custom all-reduce kernel and
            fall back to NCCL.
        tokenizer_pool_config: Config for the tokenizer pool.
            If None, will use synchronous tokenization.
        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.
    """

    def __init__(
        self,
        pipeline_parallel_size: int,
        tensor_parallel_size: int,
        worker_use_ray: bool,
        max_parallel_loading_workers: Optional[int] = None,
        disable_custom_all_reduce: bool = False,
        tokenizer_pool_config: Optional[TokenizerPoolConfig] = None,
        ray_workers_use_nsight: bool = False,
        placement_group: Optional["PlacementGroup"] = None,
    ) -> None:
        self.pipeline_parallel_size = pipeline_parallel_size
        self.tensor_parallel_size = tensor_parallel_size
        self.worker_use_ray = worker_use_ray
        self.max_parallel_loading_workers = max_parallel_loading_workers
        self.disable_custom_all_reduce = disable_custom_all_reduce
        self.tokenizer_pool_config = tokenizer_pool_config
        self.ray_workers_use_nsight = ray_workers_use_nsight
        self.placement_group = placement_group

        self.world_size = pipeline_parallel_size * self.tensor_parallel_size
        if self.world_size > 1:
            self.worker_use_ray = True
        self._verify_args()

    def _verify_args(self) -> None:
        if self.pipeline_parallel_size > 1:
            raise NotImplementedError(
                "Pipeline parallelism is not supported yet.")
        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.")
        if self.ray_workers_use_nsight and not self.worker_use_ray:
            raise ValueError("Unable to use nsight profiling unless workers "
                             "run with Ray.")


class SchedulerConfig:
    """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.
        max_model_len: Maximum length of a sequence (including prompt
            and generated text).
        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.
        delay_factor: Apply a delay (of delay factor multiplied by previous
            prompt latency) before scheduling next prompt.
        enable_chunked_prefill: If True, prefill requests can be chunked based
            on the remaining max_num_batched_tokens.
    """

    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,
    ) -> 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)
        self.max_num_seqs = max_num_seqs
        self.max_model_len = max_model_len
        self.use_v2_block_manager = use_v2_block_manager
        self.num_lookahead_slots = num_lookahead_slots
        self.delay_factor = delay_factor
        self.chunked_prefill_enabled = enable_chunked_prefill

        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}).")

        if self.num_lookahead_slots < 0:
            raise ValueError(
                "num_lookahead_slots "
                f"({self.num_lookahead_slots}) must be greater than or "
                "equal to 0.")


class DeviceConfig:

    def __init__(self, device: str = "auto") -> None:
        if device == "auto":
            # Automated device type detection
            if is_neuron():
                self.device_type = "neuron"
            elif is_cpu():
                self.device_type = "cpu"
            else:
                # We don't call torch.cuda.is_available() here to
                # avoid initializing CUDA before workers are forked
                self.device_type = "cuda"
        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)


@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 >= "
                f"max_loras ({self.max_loras})")

    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.")


@dataclass
class VisionLanguageConfig:
    """Configs the input data format and how models should run for
    vision language models."""

    class ImageInputType(enum.Enum):
        """Image input type into the vision language model.

        An image roughly goes through the following transformation:
        Raw image --> pixel values --> image features --> image embeddings.

        The difference between different image input types is where the
        image encoder (pixel values --> image features) is run.
        Different image input types also correspond to different tensor shapes.

        For example, for Llava, PIXEL_VALUES: (1, 3, 336, 336).
        IMAGE_FEATURES: (1, 576, 1024).
        """
        PIXEL_VALUES = enum.auto()
        IMAGE_FEATURES = enum.auto()

    image_input_type: ImageInputType
    # The input id corresponding to image token.
    image_token_id: int
    # Used for running `run_prefill_max_token`.
    # For models that support varying resolution, this corresponds to
    # worst case scenario (biggest supported resolution).
    image_input_shape: tuple
    image_feature_size: int

    @classmethod
    def get_image_input_enum_type(
            cls, value: str) -> "VisionLanguageConfig.ImageInputType":
        """Get the image input type from a string."""
        try:
            return cls.ImageInputType[value.upper()]
        except KeyError as e:
            raise ValueError(f"{value} is not a valid choice. "
                             f"Expecting to choose from "
                             f"{[x.name for x in cls.ImageInputType]}.") from e


_STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.float16,
    "float16": torch.float16,
    "float": torch.float32,
    "float32": torch.float32,
    "bfloat16": torch.bfloat16,
}

_ROCM_NOT_SUPPORTED_DTYPE = ["float", "float32"]


def _get_and_verify_dtype(
    config: PretrainedConfig,
    dtype: Union[str, torch.dtype],
) -> 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

    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
        else:
            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
    else:
        raise ValueError(f"Unknown dtype: {dtype}")

    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)
        ]
        raise ValueError(f"dtype '{dtype}' is not supported in ROCm. "
                         f"Supported dtypes are {rocm_supported_dtypes}")

    # 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:
            # Casting between float16 and bfloat16 is allowed with a warning.
            logger.warning(f"Casting {config_dtype} to {torch_dtype}.")

    return torch_dtype


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",
        # ChatGLM2
        "seq_length",
        # Command-R
        "model_max_length",
        # Others
        "max_sequence_length",
        "max_seq_length",
        "seq_len",
    ]
    max_len_key = None
    for key in possible_keys:
        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)
    if derived_max_model_len == float("inf"):
        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

    rope_scaling = getattr(hf_config, "rope_scaling", None)
    if rope_scaling is not None:
        assert "factor" in rope_scaling
        scaling_factor = rope_scaling["factor"]
        if rope_scaling["type"] == "yarn":
            derived_max_model_len = rope_scaling[
                "original_max_position_embeddings"]
        derived_max_model_len *= scaling_factor

    if max_model_len is None:
        max_model_len = derived_max_model_len
    elif max_model_len > derived_max_model_len:
        # 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:
            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.")
    return int(max_model_len)