99 lines
3.7 KiB
Python
99 lines
3.7 KiB
Python
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
from typing import TYPE_CHECKING, Optional
|
|
|
|
from .base import HfQuantizer
|
|
|
|
|
|
if TYPE_CHECKING:
|
|
from ..modeling_utils import PreTrainedModel
|
|
|
|
from ..utils import is_accelerate_available, is_torch_available, is_vptq_available, logging
|
|
from ..utils.quantization_config import QuantizationConfigMixin
|
|
|
|
|
|
if is_torch_available():
|
|
import torch
|
|
|
|
logger = logging.get_logger(__name__)
|
|
|
|
|
|
class VptqHfQuantizer(HfQuantizer):
|
|
"""
|
|
Quantizer of the VPTQ method. Enables the loading of prequantized models.
|
|
"""
|
|
|
|
requires_calibration = True
|
|
required_packages = ["vptq"]
|
|
|
|
def __init__(self, quantization_config: QuantizationConfigMixin, **kwargs):
|
|
super().__init__(quantization_config, **kwargs)
|
|
self.quantization_config = quantization_config
|
|
|
|
def validate_environment(self, *args, **kwargs):
|
|
if not is_accelerate_available():
|
|
raise ImportError("Using `vptq` quantization requires Accelerate: `pip install accelerate`")
|
|
|
|
if not is_vptq_available():
|
|
raise ImportError("Using `vptq` quantization requires VPTQ>=0.0.4: `pip install -U vptq`")
|
|
|
|
def update_torch_dtype(self, torch_dtype: "torch.dtype") -> "torch.dtype":
|
|
if torch_dtype is None:
|
|
if torch.cuda.is_available():
|
|
torch_dtype = torch.float16
|
|
logger.info(
|
|
"CUDA available. Assuming VPTQ inference on GPU and loading the model in `torch.float16`. To overwrite it, set `torch_dtype` manually."
|
|
)
|
|
else:
|
|
import vptq
|
|
|
|
device_availability = getattr(vptq, "device_availability", lambda device: False)
|
|
if device_availability("cpu") is True:
|
|
raise RuntimeError("No GPU found. Please wait for the next release of VPTQ to use CPU inference")
|
|
torch_dtype = torch.float32
|
|
logger.info("No GPU found. Assuming VPTQ inference on CPU and loading the model in `torch.float32`.")
|
|
return torch_dtype
|
|
|
|
def _process_model_before_weight_loading(
|
|
self,
|
|
model: "PreTrainedModel",
|
|
keep_in_fp32_modules: Optional[list[str]] = None,
|
|
**kwargs,
|
|
):
|
|
"""
|
|
we don't have param like modules_to_not_convert to indicate which layers should not be quantized
|
|
because `quantization_config` include the layers that should be quantized
|
|
"""
|
|
from ..integrations import replace_with_vptq_linear
|
|
|
|
self.modules_to_not_convert = self.get_modules_to_not_convert(
|
|
model, self.quantization_config.modules_to_not_convert, keep_in_fp32_modules
|
|
)
|
|
|
|
replace_with_vptq_linear(
|
|
model,
|
|
quantization_config=self.quantization_config,
|
|
modules_to_not_convert=self.modules_to_not_convert,
|
|
)
|
|
model.config.quantization_config = self.quantization_config
|
|
|
|
def _process_model_after_weight_loading(self, model: "PreTrainedModel", **kwargs):
|
|
return model
|
|
|
|
@property
|
|
def is_trainable(self) -> bool:
|
|
return False
|
|
|
|
def is_serializable(self, safe_serialization=None):
|
|
return True
|