mirror of
https://github.com/NVIDIA/Model-Optimizer.git
synced 2026-10-14 05:00:00 +08:00
82 lines
3.1 KiB
Python
82 lines
3.1 KiB
Python
# SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
# SPDX-License-Identifier: Apache-2.0
|
|
#
|
|
# 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.
|
|
|
|
import torch
|
|
|
|
from modelopt.torch.quantization import utils as quant_utils
|
|
from modelopt.torch.quantization.calib.max import MaxCalibrator
|
|
|
|
|
|
class PercentileCalibrator(MaxCalibrator):
|
|
def __init__(self, num_bits=8, axis=None, unsigned=False, track_amax=False, **kwargs):
|
|
super().__init__(num_bits, axis, unsigned, track_amax)
|
|
self.percentile = kwargs["percentile"]
|
|
self.total_step = kwargs["total_step"]
|
|
self.collect_method = kwargs["collect_method"]
|
|
self.data = {}
|
|
self.i = 0
|
|
|
|
def collect(self, x):
|
|
"""Tracks the absolute max of all tensors.
|
|
|
|
Args:
|
|
x: A tensor
|
|
|
|
Raises:
|
|
RuntimeError: If amax shape changes
|
|
"""
|
|
# Swap axis to reduce.
|
|
reduce_axis = quant_utils.convert_quantization_axis_to_reduce_axis(x, self._axis)
|
|
local_amax = quant_utils.reduce_amax(x, axis=reduce_axis).detach()
|
|
_cur_step = self.i % self.total_step
|
|
if _cur_step not in self.data:
|
|
self.data[_cur_step] = local_amax
|
|
elif self.collect_method == "global_min":
|
|
self.data[_cur_step] = torch.min(self.data[_cur_step], local_amax)
|
|
elif self.collect_method in {"min-max", "mean-max"}:
|
|
self.data[_cur_step] = torch.max(self.data[_cur_step], local_amax)
|
|
else:
|
|
self.data[_cur_step] += local_amax
|
|
if self._track_amax:
|
|
raise NotImplementedError
|
|
self.i += 1
|
|
|
|
def compute_amax(self):
|
|
"""Return the absolute max of all tensors collected."""
|
|
up_lim = int(self.total_step * self.percentile)
|
|
if self.collect_method == "min-mean":
|
|
amaxs_values = [self.data[i] / self.total_step for i in range(up_lim)]
|
|
else:
|
|
amaxs_values = [self.data[i] for i in range(up_lim)]
|
|
if self.collect_method == "mean-max":
|
|
act_amax = torch.vstack(amaxs_values).mean(axis=0)[0]
|
|
else:
|
|
act_amax = torch.vstack(amaxs_values).min(axis=0)[0]
|
|
self._calib_amax = act_amax
|
|
return self._calib_amax
|
|
|
|
def __str__(self):
|
|
s = "PercentileCalibrator"
|
|
return s.format(**self.__dict__)
|
|
|
|
def __repr__(self):
|
|
s = "PercentileCalibrator("
|
|
s += super(MaxCalibrator, self).__repr__()
|
|
s += " calib_amax={_calib_amax}"
|
|
if self._track_amax:
|
|
s += " amaxs={_amaxs}"
|
|
s += ")"
|
|
return s.format(**self.__dict__)
|