mindspore_gs.ptq.ptq_config 源代码

# Copyright 2024 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
"""algorithm related configs"""

from dataclasses import dataclass, field, is_dataclass, asdict
from enum import Enum
from typing import List, Union

from mindspore import dtype as msdtype
from mindspore.communication import get_group_size

from mindspore_gs.common import logger
from mindspore_gs.common.config import GSBaseConfig
from mindspore_gs.common.utils import value_check, list_value_check
from mindspore_gs.common.register import RegisterMachine
from mindspore_gs.common.gs_enum import BackendTarget

algo_cfg_register = RegisterMachine()


class PTQApproach(Enum):
    """
    PTQ approach enums
    """
    SMOOTH_QUANT = 'smooth_quant'
    RTN = 'rtn'
    GPTQ = 'gptq'
    PTQ = 'ptq'
    AWQ = 'awq'


[文档]class PTQMode(Enum): """ Mode for ptq quantizer. - ``QUANTIZE``: indicate ptq quantizer in quantize mode. - ``DEPLOY``: indicate ptq quantizer in deploy mode. """ QUANTIZE = 'quantize' DEPLOY = 'deploy'
[文档]class OutliersSuppressionType(Enum): """ Outliers suppression type for ptq quantizer. - ``SMOOTH``: apply smooth scale between weight and activate. - ``AWQ``: apply awq algo to search best smooth_scale and clip parameter. - ``NONE``: not doing any outliers suppression. """ SMOOTH = 'smooth' AWQ = 'awq' NONE = 'none' @classmethod def from_str(cls, name: str): """outlier_formatter""" if name.lower() == 'smooth': return OutliersSuppressionType.SMOOTH if name.lower() == 'awq': return OutliersSuppressionType.AWQ return OutliersSuppressionType.NONE
class LayerQuantizeAlgo(Enum): """ Quantization algorithm for each layer. - ``A16W8``: apply. """ A16W8 = 'a16w8' A8W8 = 'a8w8'
[文档]class PrecisionRecovery(Enum): """ Precision recovery algorithms. - ``GPTQ``: apply gptq algorithm to recovery precision. - ``NONE``: not doing any precision recovery. """ GPTQ = 'gptq' NONE = 'none'
[文档] @classmethod def from_str(cls, name: str): """outlier_formatter""" if name.lower() == 'gptq': return PrecisionRecovery.GPTQ return PrecisionRecovery.NONE
@algo_cfg_register.register(PTQApproach.GPTQ) @dataclass class GPTQQuantConfig: """config for gptq quant algorithm""" block_size: int = 128 desc_act: bool = False damp_percent: float = 0.01 static_groups: bool = False def __post_init__(self): value_check('block_size', self.block_size, int) value_check('desc_act', self.desc_act, bool) value_check('damp_percent', self.damp_percent, float) value_check('static_groups', self.static_groups, bool) if self.block_size < 0: raise ValueError("GPTQConfig block_size should >=0, " f"but got {self.block_size}.") if self.damp_percent < 0 or self.damp_percent > 1: raise ValueError("GPTQConfig damp_percent should >=0 and <=1, " f"but got {self.damp_percent}.") @algo_cfg_register.register(PTQApproach.SMOOTH_QUANT) @dataclass class SmoothQuantConfig: """config for smooth quant algorithm""" alpha: float = 0.5 def __post_init__(self): value_check('alpha', self.alpha, float) @algo_cfg_register.register(PTQApproach.AWQ) @dataclass class AWQConfig: """ config for awq quant algorithm - `duo_scaling`: use activation and weight to compute scale. - `smooth_alpha`: the hyper-parameter of smooth search. - `weight_clip_ratio`: the hyper-parameter of clip search. """ duo_scaling: bool = True smooth_alpha: Union[list, float] = field(default_factory=lambda: [i/20 for i in range(20)]) weight_clip_ratio: Union[list, float] = field(default_factory=lambda: [1- i/20 for i in range(10)]) def __post_init__(self): value_check('duo_scaling', self.duo_scaling, bool) if not isinstance(self.smooth_alpha, (float, list)): raise TypeError("AWQConfig smooth_alpha only support float or list, " f"but got {type(self.smooth_alpha)}") if not isinstance(self.weight_clip_ratio, (float, list)): raise TypeError("AWQConfig weight_clip_ratio only support float or list, " f"but got {type(self.weight_clip_ratio)}") if isinstance(self.smooth_alpha, float) and (self.smooth_alpha < 0 or self.smooth_alpha > 1): raise ValueError("AWQConfig smooth_alpha should >=0 and <=1, " f"but got {self.smooth_alpha}.") if isinstance(self.weight_clip_ratio, float) and (self.weight_clip_ratio < 0 or self.weight_clip_ratio > 1): raise ValueError("AWQConfig weight_clip_ratio should >=0 and <=1, " f"but got {self.weight_clip_ratio}.") if isinstance(self.smooth_alpha, list) and any(alpha < 0 or alpha > 1 for alpha in self.smooth_alpha): raise ValueError("AWQConfig smooth_alpha should >=0 and <=1, " f"but got {self.smooth_alpha}.") if isinstance(self.weight_clip_ratio, list) and any(alpha < 0 or alpha > 1 for alpha in self.weight_clip_ratio): raise ValueError("AWQConfig weight_clip_ratio should >=0 and <=1, " f"but got {self.weight_clip_ratio}.") class QuantGranularity(Enum): """quant granularity.""" PER_TENSOR = 'per_tensor' PER_CHANNEL = 'per_channel' PER_TOKEN = 'per_token' PER_GROUP = 'per_group' @classmethod def from_str(cls, name: str): """quant_granularity_formatter""" if name.lower() == 'per_token': return QuantGranularity.PER_TOKEN if name.lower() == 'per_tensor': return QuantGranularity.PER_TENSOR if name.lower() == 'per_channel': return QuantGranularity.PER_CHANNEL if name.lower() == 'per_group': return QuantGranularity.PER_GROUP return None
[文档]@dataclass class PTQConfig: """ Config for post trainning quantization. Args: mode (:class:`mindspore_gs.ptq.PTQMode`): Flag for ptq mode, ``QUANTIZATION`` for quantization mode, ``DEPLOY`` for deploy mode. backend (:class:`mindspore_gs.common.BackendTarget`): Flag for backend target, ``NONE`` for no specific backend, ``ASCEND`` for ascend backend. opname_blacklist (List[str]): Blacklist of opname. Layers in network with name fuzzy matched with this blacklist will not being quanted. algo_args (Union[dict, dataclass]): Used to configure hyperparameters of algorithms such as RTN, SmoothQuant, and OmniQuant. act_quant_dtype (mindspore.dtype): Used to configure the quantization type of activation. mindspore.dtype.int8 indicates that the activation is quantized by 8 bits, and None indicates that it is not quantized. weight_quant_dtype (mindspore.dtype): Used to configure the quantization type of weight. mindspore.dtype.int8 indicates that the weight is quantized by 8 bits, and None indicates that it is not quantized. kvcache_quant_dtype (mindspore.dtype): Used to configure the quantization type of kvcache. mindspore.dtype.int8 indicates that the kvcache is quantized by 8 bits, and None indicates that it is not quantized. outliers_suppression (:class:`mindspore_gs.ptq.OutliersSuppressionType`): Used to configure outliers suppression method before quantization. OutliersSuppressionType.SMOOTH indicates using smooth method from SmoothQuant to suppress outliers, and OutliersSuppressionType.NONE as default indicates doing nothing for outliers. precision_recovery (:class:`mindspore_gs.ptq.PrecisionRecovery`): Used to precision compensation of weights during quantization. PrecisionRecovery.GPTQ indicates using GPTQ method to compensate precision, and PrecisionRecovery.NONE as default indicates doing nothing for precision recovery. act_quant_granularity (:class:`mindspore_gs.ptq.QuantGranularity`): Used to configure the quantization granularity of activation. Currently only QuantGranularity.PER_TENSOR and QuantGranularity.PER_TOKEN are supported. kvcache_quant_granularity (:class:`mindspore_gs.ptq.QuantGranularity`): Used to configure the quantization granularity of kvcache. Currently only QuantGranularity.PER_CHANNEL and QuantGranularity.PER_TOKEN are supported. weight_quant_granularity (:class:`mindspore_gs.ptq.QuantGranularity`): Used to configure the quantization granularity of weight. Currently only QuantGranularity.PER_CHANNEL and QuantGranularity.PER_GROUP are supported. group_size (int, optional): group_size of per_group quantization, suggest using 64 or 128. Default value: ``0``. Raises: ValueError: If `mode` is not PTQMode.QUANTIZE or PTQMode.DEPLOY. ValueError: If `backend` is not BackendTarget.NONE or BackendTarget.ASCEND. TypeError: If `opname_blacklist` is not a list of str. ValueError: If `weight_quant_dtype` is not mindspore.dtype.int8 or None. ValueError: If `kvcache_quant_dtype` is not mindspore.dtype.int8 or None. ValueError: If `act_quant_dtype` is not mindspore.dtype.int8 or None. TypeError: If `outliers_suppression` is not a OutliersSuppressionType. TypeError: If `precision_recovery` is not a PrecisionRecovery. ValueError: If `act_quant_granularity` is not QuantGranularity.PER_TENSOR or QuantGranularity.PER_TOKEN. ValueError: If `kvcache_quant_granularity` is not QuantGranularity.PER_CHANNEL or QuantGranularity.PER_TOKEN. ValueError: If `act_quant_granularity` is QuantGranularity.PER_TOKEN but weight_quant_dtype != msdtype.int8 or act_quant_dtype != msdtype.int8. ValueError: If `kvcache_quant_granularity` is QuantGranularity.PER_TOKEN but kvcache_quant_dtype != msdtype.int8. ValueError: If `weight_quant_granularity` is not QuantGranularity.PER_CHANNEL or QuantGranularity.PER_GROUP. ValueError: If `weight_quant_granularity` is QuantGranularity.PER_GROUP but `group_size` is not in [64, 128]. ValueError: If `weight_quant_granularity` is not QuantGranularity.PER_GROUP but `group_size` != 0. TypeError: If `group_size` is not Int. Examples: >>> from mindspore_gs.ptq import PTQConfig, PTQMode >>> from mindspore_gs.common import BackendTarget >>> PTQConfig(mode=PTQMode.DEPLOY, backend=BackendTarget.ASCEND, opname_blacklist=['layer0']) PTQConfig(mode=<PTQMode.DEPLOY: 'deploy'>, backend=<BackendTarget.ASCEND: 'ascend'>, opname_blacklist=['layer0'], algo_args={}) """ mode: PTQMode = PTQMode.QUANTIZE backend: BackendTarget = BackendTarget.ASCEND opname_blacklist: List[str] = field(default_factory=list) algo_args: Union[dict, dataclass] = field(default_factory=dict) weight_quant_dtype: msdtype = msdtype.int8 kvcache_quant_dtype: msdtype = None act_quant_dtype: msdtype = None precision_recovery: PrecisionRecovery = PrecisionRecovery.NONE outliers_suppression: OutliersSuppressionType = OutliersSuppressionType.NONE weight_quant_granularity: QuantGranularity = QuantGranularity.PER_CHANNEL kvcache_quant_granularity: QuantGranularity = QuantGranularity.PER_CHANNEL act_quant_granularity: QuantGranularity = QuantGranularity.PER_TENSOR group_size: int = 0 def __post_init__(self): weight_support = [msdtype.int8, msdtype.qint4x2, None] act_support = [msdtype.int8, None] kvcache_support = [msdtype.int8, None] if self.mode not in PTQMode.__members__.values(): raise ValueError(f'mode shall be in {PTQMode.__members__.values()}') if self.backend not in BackendTarget.__members__.values(): raise ValueError(f'backend shall be in {BackendTarget.__members__.values()}') if self.weight_quant_dtype not in weight_support: raise ValueError(f'weight_quant_dtype support {weight_support}, but got {self.weight_quant_dtype}.') if self.kvcache_quant_dtype not in kvcache_support: raise ValueError(f'kvcache_quant_dtype support {kvcache_support}, but got {self.kvcache_quant_dtype}.') if self.act_quant_dtype not in act_support: raise ValueError(f'act_quant_dtype support {act_support}, but got {self.act_quant_dtype}.') list_value_check('opname_blacklist', self.opname_blacklist, str) value_check('outliers_suppression', self.outliers_suppression, OutliersSuppressionType) value_check('group_size', self.group_size, int) self._check_quant_granularity() self._check_precision_recovery() if not isinstance(self.algo_args, dict) and not is_dataclass(self.algo_args): raise ValueError(f"algo_args's type should be dict or dataclass, but now is {type(self.algo_args)}") if self.algo_args and is_dataclass(self.algo_args): self.algo_args = asdict(self.algo_args) def _check_precision_recovery(self): '''check precision recovery''' if self.precision_recovery == PrecisionRecovery.GPTQ and self.algo_args == {}: self.algo_args = GPTQQuantConfig() logger.warning('GPTQQuantConfig is not configured, it will apply the default parameters.') if self.precision_recovery != PrecisionRecovery.GPTQ and isinstance(self.algo_args, GPTQQuantConfig): logger.warning(f'GPTQQuantConfig is configured, but the precision recovery is {self.precision_recovery}.') if self.precision_recovery == PrecisionRecovery.GPTQ and \ self.weight_quant_granularity == QuantGranularity.PER_GROUP and \ (self.algo_args.desc_act and not self.algo_args.static_groups): raise ValueError('when using GPTQ algorithm with per_group quantization, ' 'if desc_act is True ,then static_groups must be true.') value_check('precision_recovery', self.precision_recovery, PrecisionRecovery) def _check_quant_granularity(self): """check_quant_granularity""" act_quant_granularity_support = [QuantGranularity.PER_TENSOR, QuantGranularity.PER_TOKEN] weight_quant_granularity_support = [QuantGranularity.PER_CHANNEL, QuantGranularity.PER_GROUP] kvcache_quant_granularity_support = [QuantGranularity.PER_CHANNEL, QuantGranularity.PER_TOKEN] if self.act_quant_granularity not in act_quant_granularity_support: raise ValueError(f'act_quant_granularity support {act_quant_granularity_support}, ' f'but got {self.act_quant_granularity}.') if self.weight_quant_granularity not in weight_quant_granularity_support: raise ValueError(f'weight_quant_granularity support {weight_quant_granularity_support}, ' f'but got {self.weight_quant_granularity}.') if self.kvcache_quant_granularity not in kvcache_quant_granularity_support: raise ValueError(f'kvcache_quant_granularity support {kvcache_quant_granularity_support}, ' f'but got {self.kvcache_quant_granularity}.') if (self.weight_quant_dtype != msdtype.int8 or self.act_quant_dtype != msdtype.int8) and \ self.act_quant_granularity is QuantGranularity.PER_TOKEN: raise ValueError(f'when act_quant_granularity is QuantGranularity.PER_TOKEN, weight_quant_dtype: {self.weight_quant_dtype} ' f'and act_quant_dtype: {self.act_quant_dtype} must be mindspore.dtype.int8.') if (self.weight_quant_dtype == msdtype.int8 and self.act_quant_dtype == msdtype.int8) and \ self.weight_quant_granularity != QuantGranularity.PER_CHANNEL: raise ValueError(f'when weight_quant_dtype is int8 and act_quant_dtype is int8, the weight_quant_granularity must be QuantGranularity.PER_CHANNEL,' f' but got {self.weight_quant_granularity}.') if self.kvcache_quant_dtype != msdtype.int8 and self.kvcache_quant_granularity is QuantGranularity.PER_TOKEN: raise ValueError('when kvcache_quant_granularity is QuantGranularity.PER_TOKEN, kvcache_quant_dtype must be mindspore.dtype.int8.') if self.mode == PTQMode.QUANTIZE and self.kvcache_quant_granularity is QuantGranularity.PER_TOKEN: logger.warning('kvcache_quant_granularity is QuantGranularity.PER_TOKEN, not need quantize for kvcache.') if self.weight_quant_granularity != QuantGranularity.PER_GROUP and self.group_size != 0: raise ValueError("group_size should equal to 0 when not to do pre_group quantize.") if self.weight_quant_granularity == QuantGranularity.PER_GROUP and self.group_size not in [64, 128]: raise ValueError("group_size should be in [64, 128] when doing pre_group quantize.")
class YamlLoader: """Loader for some special item in yaml.""" def __call__(self, src: str): raise NotImplementedError class MSDTypeLoader(YamlLoader): """Loader for `mindspore.dtype` in yaml.""" def __init__(self): self.dtype_dict = { "Bool": msdtype.bool_, "Int": msdtype.int_, "Int4": msdtype.qint4x2, "Int8": msdtype.int8, "Int16": msdtype.int16, "Int32": msdtype.int32, "Int64": msdtype.int64, "UInt8": msdtype.uint8, "UInt16": msdtype.uint16, "UInt32": msdtype.uint32, "UInt64": msdtype.uint64, "Float": msdtype.float_, "Float16": msdtype.float16, "Float32": msdtype.float32, "Float64": msdtype.float64, "BFloat16": msdtype.bfloat16, "Complex64": msdtype.complex64, "Complex128": msdtype.complex128, } def __call__(self, src: str): if src == "None": return None ms_dtype = self.dtype_dict.get(src, None) if not ms_dtype: raise ValueError(f"Unrecognized dtype: {src}") return ms_dtype @dataclass class InnerPTQConfig(GSBaseConfig, PTQConfig): """ config for post-trainning-quantizer """ approach: PTQApproach = field(default=PTQApproach.RTN) act_per_channel: bool = False weight_per_channel: bool = True kvcache_per_head: bool = True act_symmetric: bool = False weight_symmetric: bool = True kvcache_symmetric: bool = True act_narrow_range: bool = False weight_narrow_range: bool = False kvcache_narrow_range: bool = False enable_deploy_fusion: bool = True kvcache_calibrate_max_new_tokens: int = 10 reflash_inputs_after_each_processor: bool = False fallback_blacklist: dict = field(default_factory=dict) tp_size: int = 1 def update_tp_size(self): try: self.tp_size = get_group_size() except RuntimeError: pass def __post_init__(self): self.update_tp_size() value_check('act_per_channel', self.act_per_channel, bool) value_check('weight_per_channel', self.weight_per_channel, bool) value_check('kvcache_per_head', self.kvcache_per_head, bool) value_check('act_symmetric', self.act_symmetric, bool) value_check('weight_symmetric', self.weight_symmetric, bool) value_check('kvcache_symmetric', self.kvcache_symmetric, bool) value_check('act_narrow_range', self.act_narrow_range, bool) value_check('weight_narrow_range', self.weight_narrow_range, bool) value_check('enable_deploy_fusion', self.enable_deploy_fusion, bool) value_check('kvcache_calibrate_max_new_tokens', self.kvcache_calibrate_max_new_tokens, int) value_check('fallback_blacklist', self.fallback_blacklist, dict) value_check('reflash_inputs_after_each_processor', self.reflash_inputs_after_each_processor, bool) if self.approach not in PTQApproach.__members__.values(): raise ValueError(f'Invalid approach: {self.approach}') self._check_quant_granularity() self._check_rtn() if list(set(self.fallback_blacklist.keys()) & set(self.opname_blacklist)): raise ValueError("There should be no repetition between opname_blacklist and fallback_a16w8_blacklist," f"now opname_blacklist={self.opname_blacklist}," f"fallback_a16w8_blacklist={self.fallback_blacklist}") if not self.algo_args: args_config = algo_cfg_register[self.approach] if args_config is not None and is_dataclass(args_config): self.algo_args.update(asdict(args_config())) def _check_rtn(self): """_check_rtn""" if self.approach is PTQApproach.RTN and self.act_quant_dtype == msdtype.int8 and self.act_quant_granularity is not QuantGranularity.PER_TOKEN: raise ValueError(f"{self.approach} is not support act_quant_dtype == mindspore.dtype.int8 when act_quant_granularity is not QuantGranularity.PER_TOKEN.") if self.approach is PTQApproach.RTN and self.weight_quant_dtype is None and self.kvcache_quant_dtype is None: raise ValueError(f"weight_quant_dtype and kvcache_quant_dtype are None, {self.approach} can't take effect.") if self.approach is PTQApproach.RTN and self.weight_quant_dtype == msdtype.int8 and self.kvcache_quant_dtype == msdtype.int8 \ and self.kvcache_quant_granularity is not QuantGranularity.PER_TOKEN: raise ValueError("when self.kvcache_quant_granularity not QuantGranularity.PER_TOKEN, weight_quant_dtype and kvcache_quant_dtype are mindspore.dtype.int8, " f"{self.approach} isn't supported.") if self.approach is PTQApproach.RTN and self.kvcache_quant_dtype == msdtype.int8 and self.kvcache_quant_granularity is not QuantGranularity.PER_TOKEN: raise ValueError("The c8 quant in RTN is discarded.") def _check_quant_granularity(self): if self.approach is PTQApproach.RTN and (self.act_quant_granularity is QuantGranularity.PER_TOKEN or \ self.kvcache_quant_granularity is QuantGranularity.PER_TOKEN) and self.mode is PTQMode.QUANTIZE: raise ValueError("self.mode is PTQMode.QUANTIZE, self.act_quant_granularity is QuantGranularity.PER_TOKEN or " f"self.kvcache_quant_granularity is QuantGranularity.PER_TOKEN, {self.approach} can't take effect.") if (self.act_quant_granularity is QuantGranularity.PER_TOKEN or self.kvcache_quant_granularity is QuantGranularity.PER_TOKEN) and \ self.approach is not PTQApproach.RTN and self.approach is not PTQApproach.PTQ: raise ValueError("self.act_quant_granularity is QuantGranularity.PER_TOKEN or " f"self.kvcache_quant_granularity is QuantGranularity.PER_TOKEN, {self.approach} can't take effect.") if self.act_quant_granularity is QuantGranularity.PER_TOKEN and self.weight_symmetric is False: raise ValueError("when self.act_quant_granularity is QuantGranularity.PER_TOKEN, self.weight_symmetric must be True.") def _parse_dict(self): """ parse data class to readable dicts""" parsed_dict = self.__dict__ parsed_dict['backend'] = self.backend.name parsed_dict['mode'] = self.mode.name parsed_dict['approach'] = self.approach.name parsed_dict['opname_blacklist'] = self.opname_blacklist parsed_dict['kvcache_quant_dtype'] = str(self.kvcache_quant_dtype) parsed_dict['weight_quant_dtype'] = str(self.weight_quant_dtype) parsed_dict['act_quant_dtype'] = str(self.act_quant_dtype) parsed_dict['precision_recovery'] = self.precision_recovery.name parsed_dict['outliers_suppression'] = self.outliers_suppression.name parsed_dict['act_quant_granularity'] = self.act_quant_granularity.name parsed_dict['kvcache_quant_granularity'] = self.kvcache_quant_granularity.name parsed_dict['weight_quant_granularity'] = self.weight_quant_granularity.name return parsed_dict def _unparse_dict(self, data_dict): """ convert readable dicts to data config""" def update_dict(key, decode_fn): nonlocal data_dict if key not in data_dict: raise ValueError(f'{key} shall in yaml, but not found') if isinstance(decode_fn, YamlLoader): data_dict[key] = decode_fn(data_dict[key]) else: data_dict[key] = decode_fn[data_dict[key]] unparse_list = [ ('mode', PTQMode), ('backend', BackendTarget), ('approach', PTQApproach), ('outliers_suppression', OutliersSuppressionType), ('precision_recovery', PrecisionRecovery), ('kvcache_quant_dtype', MSDTypeLoader()), ('weight_quant_dtype', MSDTypeLoader()), ('act_quant_dtype', MSDTypeLoader()), ('act_quant_granularity', QuantGranularity), ('kvcache_quant_granularity', QuantGranularity), ('weight_quant_granularity', QuantGranularity), ] for item in unparse_list: update_dict(*item) self.__dict__.update(data_dict) @staticmethod def inner_config(cfg: PTQConfig, approach=None): """convert PTQConfig to InnerConfig""" if not isinstance(cfg, PTQConfig): raise TypeError(f'input config shall be PTQConfig, but got {type(cfg)}') if not approach: inner_cfg = InnerPTQConfig() else: inner_cfg = InnerPTQConfig(approach=approach) for key, val in asdict(cfg).items(): if key == "algo_args": inner_cfg.algo_args.update(val) else: setattr(inner_cfg, key, val) return inner_cfg