ChatTTS/model/velocity/model_loader.py

"""Utilities for selecting and loading models."""

import contextlib

import torch
import torch.nn as nn

from vllm.config import ModelConfig
from vllm.model_executor.models import ModelRegistry
from vllm.model_executor.weight_utils import get_quant_config, initialize_dummy_weights

from .llama import LlamaModel


@contextlib.contextmanager
def _set_default_torch_dtype(dtype: torch.dtype):
    """Sets the default torch dtype to the given dtype."""
    old_dtype = torch.get_default_dtype()
    torch.set_default_dtype(dtype)
    yield
    torch.set_default_dtype(old_dtype)


def get_model(model_config: ModelConfig) -> nn.Module:
    # Get the (maybe quantized) linear method.
    linear_method = None
    if model_config.quantization is not None:
        quant_config = get_quant_config(
            model_config.quantization,
            model_config.model,
            model_config.hf_config,
            model_config.download_dir,
        )
        capability = torch.cuda.get_device_capability()
        capability = capability[0] * 10 + capability[1]
        if capability < quant_config.get_min_capability():
            raise ValueError(
                f"The quantization method {model_config.quantization} is not "
                "supported for the current GPU. "
                f"Minimum capability: {quant_config.get_min_capability()}. "
                f"Current capability: {capability}."
            )
        supported_dtypes = quant_config.get_supported_act_dtypes()
        if model_config.dtype not in supported_dtypes:
            raise ValueError(
                f"{model_config.dtype} is not supported for quantization "
                f"method {model_config.quantization}. Supported dtypes: "
                f"{supported_dtypes}"
            )
        linear_method = quant_config.get_linear_method()

    with _set_default_torch_dtype(model_config.dtype):
        # Create a model instance.
        # The weights will be initialized as empty tensors.
        with torch.device("cuda"):
            model = LlamaModel(model_config.hf_config, linear_method)
        if model_config.load_format == "dummy":
            # NOTE(woosuk): For accurate performance evaluation, we assign
            # random values to the weights.
            initialize_dummy_weights(model)
        else:
            # Load the weights from the cached or downloaded files.
            model.load_weights(
                model_config.model,
                model_config.download_dir,
                model_config.load_format,
                model_config.revision,
            )
    return model.eval()