image-match/python/py/Lib/site-packages/transformers/quantizers/quantizer_fouroversix.py

from typing import TYPE_CHECKING

from ..utils.import_utils import is_fouroversix_available
from .base import HfQuantizer
from .quantizers_utils import get_module_from_name


if TYPE_CHECKING:
    from ..modeling_utils import PreTrainedModel
    from ..utils.quantization_config import FourOverSixConfig

from ..utils import (
    is_torch_available,
)


if is_torch_available():
    import torch


class FourOverSixHfQuantizer(HfQuantizer):
    """
    FP4 quantization with fouroversix.
    """

    requires_calibration = False
    quantization_config: "FourOverSixConfig"

    def __init__(self, quantization_config, **kwargs):
        super().__init__(quantization_config, **kwargs)

    def validate_environment(self, *args, **kwargs):
        if not is_fouroversix_available():
            raise ImportError(
                "Using `fouroversix` requires fouroversix: `pip install fouroversix --no-build-isolation`"
            )

    def param_element_size(
        self,
        model: "PreTrainedModel",
        param_name: str,
        param: "torch.Tensor",
    ) -> float:
        from fouroversix import QuantizedModule

        module, tensor_name = get_module_from_name(model, param_name)

        if QuantizedModule.is_quantized_module_type(type(module)):
            return module.get_element_size(tensor_name)

        return super().param_element_size(model, param_name, param)

    def param_needs_quantization(
        self,
        model: "PreTrainedModel",
        param_name: str,
        **kwargs,
    ) -> bool:
        from fouroversix import QuantizedModule

        module, tensor_name = get_module_from_name(model, param_name)

        return QuantizedModule.is_quantized_module_type(type(module)) and tensor_name in module.parameters_to_quantize

    def _process_model_before_weight_loading(
        self,
        model: "PreTrainedModel",
        device_map,
        **kwargs,
    ):
        from fouroversix import QuantizedModule, quantize_model

        from ..integrations.fouroversix import adapt_fouroversix_config

        quantize_model(
            model,
            adapt_fouroversix_config(self.quantization_config),
        )

        # If the model has already been quantized, we need to delete the weight tensor here so that
        # it's not expected when parameters are loaded from the checkpoint.
        if self.pre_quantized and not self.quantization_config.keep_master_weights:
            for _, module in model.named_modules():
                if QuantizedModule.is_quantized_module_type(type(module)):
                    for parameter_name in module.parameters_to_quantize:
                        delattr(module, parameter_name)

    def _process_model_after_weight_loading(self, model: "PreTrainedModel", **kwargs):
        return model

    def is_serializable(self):
        return True

    @property
    def is_trainable(self) -> bool:
        return self.quantization_config.keep_master_weights

    def get_quantize_ops(self):
        from ..integrations.fouroversix import FourOverSixQuantize

        return FourOverSixQuantize(self)

    def get_weight_conversions(self):
        """
        Return weight conversions for loading pre-quantized checkpoints of
        other pre-quantized models (not fouroversix models). After first use,
        the pre_quantized_model_config_type attribute is set to None to ensure
        subsequent calls (e.g., during save_pretrained) return an empty list
        since, by then, the model will be saved with our framework's format
        so weight conversions are no longer needed.
        """
        from fouroversix import WeightConversions

        # pre_quantized_model_config_type is only set if we are loading a
        # pre-quantized model so it is not guaranteed to exist.
        if hasattr(self.quantization_config, "pre_quantized_model_config_type"):
            model_config_type = self.quantization_config.pre_quantized_model_config_type
            weight_conversions = WeightConversions.get_weight_conversions(
                model_config_type,
            )
            return weight_conversions

        return []