You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

326 lines
12 KiB

4 days ago
# 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 abc import ABC, abstractmethod
from typing import TYPE_CHECKING, Any
from ..utils import is_torch_available, logging
from ..utils.quantization_config import QuantizationConfigMixin, QuantizationMethod
from .quantizers_utils import get_module_from_name
if TYPE_CHECKING:
from ..modeling_utils import PreTrainedModel
if is_torch_available():
import torch
from torch.nn import ModuleList
else:
ModuleList = str
logger = logging.get_logger(__file__)
def get_keys_to_not_convert(model) -> list:
r"""
Function to automatically detect keys to not convert for usage like quantization. For example for CausalLM modules
we may want to keep the lm_head in full precision for numerical stability reasons.
"""
# remove tied weights
tied_keys = set()
if len(model.all_tied_weights_keys) > 0:
tied_keys = set(model.all_tied_weights_keys.values()) | set(model.all_tied_weights_keys.keys())
# remove last module
last_module_key = {list(model.named_parameters())[-1][0]}
# remove output emb
output_emb_module = model.get_output_embeddings()
output_emb_keys = {
name
for name, module in model.named_modules()
if output_emb_module is not None and id(module) == id(output_emb_module)
}
modules_to_not_convert = tied_keys | last_module_key | output_emb_keys
modules_to_not_convert = list({k.removesuffix(".weight") for k in modules_to_not_convert})
return list(modules_to_not_convert)
def _assign_is_quantized(model):
from ..modeling_utils import PreTrainedModel
for module in model.modules():
if isinstance(module, PreTrainedModel):
module.config._is_quantized = True
class HfQuantizer(ABC):
"""
Abstract class of the HuggingFace quantizer. Supports for now quantizing HF transformers models for inference and/or quantization.
This class is used only for transformers.PreTrainedModel.from_pretrained and cannot be easily used outside the scope of that method
yet.
Attributes
quantization_config (`transformers.utils.quantization_config.QuantizationConfigMixin`):
The quantization config that defines the quantization parameters of your model that you want to quantize.
requires_calibration (`bool`):
Whether the quantization method requires to calibrate the model before using it.
"""
requires_calibration = False
def __init__(self, quantization_config: QuantizationConfigMixin, **kwargs):
self.quantization_config = quantization_config
self.pre_quantized = kwargs.pop("pre_quantized", True)
if not self.pre_quantized and self.requires_calibration:
raise ValueError(
f"The quantization method {quantization_config.quant_method} does require the model to be pre-quantized."
f" You explicitly passed `pre_quantized=False` meaning your model weights are not quantized. Make sure to "
f"pass `pre_quantized=True` while knowing what you are doing."
)
def update_dtype(self, dtype: "torch.dtype") -> "torch.dtype":
"""
Some quantization methods require to explicitly set the dtype of the model to a
target dtype. You need to override this method in case you want to make sure that behavior is
preserved
Args:
dtype (`torch.dtype`):
The input dtype that is passed in `from_pretrained`
"""
return dtype
def update_device_map(self, device_map: dict[str, Any] | None) -> dict[str, Any] | None:
"""
Override this method if you want to pass a override the existing device map with a new
one. E.g. for bitsandbytes, since `accelerate` is a hard requirement, if no device_map is
passed, the device_map is set to `"auto"``
Args:
device_map (`Union[dict, str]`, *optional*):
The device_map that is passed through the `from_pretrained` method.
"""
return device_map
def param_element_size(self, model: "PreTrainedModel", param_name: str, param: "torch.Tensor") -> float:
return param.element_size()
def adjust_max_memory(self, max_memory: dict[str, int | str]) -> dict[str, int | str]:
"""adjust max_memory argument for infer_auto_device_map() if extra memory is needed for quantization"""
return max_memory
def param_needs_quantization(self, model: "PreTrainedModel", param_name: str, **kwargs) -> bool:
"""
Check whether a given param needs to be quantized.
"""
return False
def validate_environment(self, *args, **kwargs):
"""
This method is used to potentially check for potential conflicts with arguments that are
passed in `from_pretrained`. You need to define it for all future quantizers that are integrated with transformers.
If no explicit check are needed, simply return nothing.
"""
return
def update_tp_plan(self, config):
"updates the tp plan for the scales"
return config
def update_ep_plan(self, config):
"updates the tp plan for the scales"
return config
def _process_model_before_weight_loading(self, model, **kwargs):
return model
def preprocess_model(self, model: "PreTrainedModel", dtype=None, **kwargs):
"""
Setting model attributes and/or converting model before weights loading. At this point
the model should be initialized on the meta device so you can freely manipulate the skeleton
of the model in order to replace modules in-place. Make sure to override the abstract method `_process_model_before_weight_loading`.
Args:
model (`~transformers.PreTrainedModel`):
The model to quantize
kwargs (`dict`, *optional*):
The keyword arguments that are passed along `_process_model_before_weight_loading`.
"""
model.is_quantized = True
model.quantization_method = self.quantization_config.quant_method
if self.pre_quantized:
self._convert_model_for_quantization(model)
self._process_model_before_weight_loading(model, **kwargs)
def _process_model_after_weight_loading(self, model: "PreTrainedModel", **kwargs):
return model
def postprocess_model(self, model: "PreTrainedModel", **kwargs):
"""
Post-process the model post weights loading.
Make sure to override the abstract method `_process_model_after_weight_loading`.
Args:
model (`~transformers.PreTrainedModel`):
The model to quantize
kwargs (`dict`, *optional*):
The keyword arguments that are passed along `_process_model_after_weight_loading`.
"""
model.config.quantization_config = self.quantization_config
if self.pre_quantized and getattr(self.quantization_config, "dequantize", False):
self.remove_quantization_config(model)
else:
_assign_is_quantized(model)
return self._process_model_after_weight_loading(model, **kwargs)
def remove_quantization_config(self, model):
"""
Remove the quantization config from the model.
"""
if hasattr(model, "hf_quantizer"):
del model.hf_quantizer
if hasattr(model.config, "quantization_config"):
del model.config.quantization_config
if hasattr(model, "quantization_method"):
del model.quantization_method
model.is_quantized = False
def dequantize(self, model, dtype=None):
"""
Potentially dequantize the model to retrieve the original model, with some loss in accuracy / performance.
Note not all quantization schemes support this.
"""
if dtype is None:
# using the same dtype we used to load the model. If we don't do that, we might have issues with modules we didn't quantize.
# or we need to upcast everything to the same dtype
dtype = model.config.dtype
model = self._dequantize(model, dtype=dtype)
self.remove_quantization_config(model)
return model
def _dequantize(self, model, dtype=None):
raise NotImplementedError(
f"{self.quantization_config.quant_method} has no implementation of `dequantize`, please raise an issue on GitHub."
)
def get_param_name(self, param_name: str) -> str:
"""
Override this method if you want to adjust the `param_name`.
"""
return param_name
@staticmethod
def get_modules_to_not_convert(
model: "PreTrainedModel",
skip_modules: list[str] | None = None,
keep_in_fp32_modules: list[str] | None = None,
add_default_skips: bool = False,
):
if skip_modules is None or add_default_skips:
modules_to_not_convert = get_keys_to_not_convert(model)
else:
modules_to_not_convert = []
if skip_modules is not None:
modules_to_not_convert.extend(skip_modules)
if keep_in_fp32_modules is not None:
modules_to_not_convert.extend(keep_in_fp32_modules)
modules_to_not_convert = list(set(modules_to_not_convert))
return modules_to_not_convert
@property
def is_qat_trainable(self) -> bool:
"""Flag indicating whether the quantized model can carry out quantization aware training"""
return False
@property
def is_compileable(self) -> bool:
"""Flag indicating whether the quantized model can be compiled"""
return False
def get_state_dict_and_metadata(self, model):
"""Get state dict and metadata. Useful when we need to modify a bit the state dict due to quantization"""
return None, {}
@abstractmethod
def is_serializable(self): ...
@property
@abstractmethod
def is_trainable(self): ...
def _convert_model_for_quantization(self, model):
for name, module in model.named_modules():
module_class_name = module.__class__.__name__
if module_class_name in MODULES_TO_PATCH_FOR_QUANTIZATION and (
self.quantization_config.quant_method
in MODULES_TO_PATCH_FOR_QUANTIZATION[module_class_name]["quantization_methods"]
):
with torch.device("meta"):
parent_module, name = get_module_from_name(model, name)
parent_module._modules[name] = MODULES_TO_PATCH_FOR_QUANTIZATION[module_class_name]["module_name"](
model.config.get_text_config()
)
def get_quantize_ops(self):
raise NotImplementedError(
f"{self.quantization_config.quant_method} is not available yet and will be supported soon."
)
def get_weight_conversions(self):
return []
class SequentialLlama4TextExperts(ModuleList):
"""
A module that implements a compressed version of a list of expert modules.
This is specifically designed to work with Llama4TextExperts in MoE layers.
"""
def __init__(self, config):
from transformers.models.llama4.modeling_llama4 import Llama4TextMLP
super().__init__([Llama4TextMLP(config) for _ in range(config.num_local_experts)])
self.num_experts = config.num_local_experts
def forward(
self,
hidden_states: "torch.Tensor",
) -> "torch.Tensor":
hidden_states = hidden_states.reshape(self.num_experts, -1, hidden_states.shape[-1])
routed_out = torch.zeros_like(hidden_states)
for expert_idx in range(self.num_experts):
routed_out[expert_idx] = self[expert_idx](hidden_states[expert_idx])
return routed_out
MODULES_TO_PATCH_FOR_QUANTIZATION = {
"Llama4TextExperts": {
"module_name": SequentialLlama4TextExperts,
"quantization_methods": [
QuantizationMethod.COMPRESSED_TENSORS,
QuantizationMethod.BITS_AND_BYTES,
],
}
}