-
Notifications
You must be signed in to change notification settings - Fork 60
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
- Loading branch information
Showing
3 changed files
with
141 additions
and
0 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,88 @@ | ||
# Copyright 2024 The HuggingFace 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 Optional, Tuple, Union | ||
|
||
import torch | ||
|
||
from ..qtype import qint2, qint4 | ||
from ..quantizers import AffineQuantizer | ||
from .max_optimizer import MaxOptimizer | ||
|
||
|
||
__all__ = ["HqqOptimizer"] | ||
|
||
|
||
# Shrinking operator | ||
def shrink_lp_op(x: torch.Tensor, beta: float, lp_norm: float) -> torch.Tensor: | ||
if lp_norm == 1: | ||
return torch.sign(x) * torch.nn.functional.relu(torch.abs(x) - 1.0 / beta) | ||
else: | ||
return torch.sign(x) * torch.nn.functional.relu( | ||
torch.abs(x) - (1.0 / beta) * torch.pow(torch.abs(x), lp_norm - 1) | ||
) | ||
|
||
|
||
class HqqOptimizer(MaxOptimizer): | ||
"""Implementation of the HQQ algorithm | ||
This is an implementation of the algorithm described in "Half-Quadratic Quantization of Large Machine Learning Models", | ||
by Hicham Badri and Appu Shaji (https://mobiusml.github.io/hqq_blog/). | ||
This is an adaption of the original implementation at https://github.com/mobiusml/hqq. | ||
""" | ||
|
||
def __init__( | ||
self, | ||
lp_norm: Optional[float] = 0.7, | ||
beta: Optional[int] = 1e1, | ||
kappa: Optional[float] = 1.01, | ||
iters: Optional[int] = 20, | ||
verbose: Optional[bool] = False, | ||
) -> None: | ||
self.lp_norm = lp_norm | ||
self.beta = beta | ||
self.kappa = kappa | ||
self.iters = iters | ||
self.verbose = verbose | ||
|
||
def optimize( | ||
self, base: torch.Tensor, bits: int, axis: int | ||
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]: | ||
scale, shift = super().optimize(base, bits, axis) | ||
best_error = None | ||
beta = self.beta | ||
qtype = qint2 if bits == 2 else qint4 | ||
base_q = AffineQuantizer.apply(base, qtype, axis, None, scale, shift) | ||
for i in range(self.iters): | ||
error = base - base_q | ||
if best_error is None: | ||
best_error = float(torch.abs(base - base_q).mean()) | ||
if self.verbose: | ||
print(f"Start error: {best_error:.6f}") | ||
e = shrink_lp_op(error, beta, self.lp_norm) | ||
mean_axis = 0 if axis == -1 else -1 | ||
hqq_shift = torch.mean(base_q._data * scale - (base - e), axis=mean_axis, keepdim=True) | ||
base_q = AffineQuantizer.apply(base, qtype, axis, None, scale, hqq_shift) | ||
mean_error = float(torch.abs(base - base_q).mean()) | ||
if self.verbose: | ||
print(f"HQQ error at it #{i}: {mean_error:.6f}") | ||
if mean_error < best_error: | ||
best_error = mean_error | ||
shift = hqq_shift | ||
beta *= self.kappa | ||
else: | ||
break | ||
|
||
return scale, shift |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,52 @@ | ||
# Copyright 2024 The HuggingFace 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. | ||
|
||
import pytest | ||
import torch | ||
from helpers import random_tensor | ||
|
||
from optimum.quanto import ( | ||
AffineQuantizer, | ||
HqqOptimizer, | ||
MaxOptimizer, | ||
qint2, | ||
qint4, | ||
) | ||
|
||
|
||
def compare_quantized_tensor(a, qtype, axis, group_size, scale, shift): | ||
qa = AffineQuantizer.apply(a, qtype, axis, group_size, scale, shift) | ||
# Evaluate mean absolute error | ||
mean_error = torch.mean(torch.abs(a - qa)) | ||
# Also evaluate cosine similarity | ||
sim = torch.nn.functional.cosine_similarity(a.flatten(), qa.flatten(), dim=0) | ||
return mean_error, sim | ||
|
||
|
||
@pytest.mark.parametrize("input_shape", [(1024, 1024), (1024, 10, 1024)]) | ||
@pytest.mark.parametrize("dtype", [torch.bfloat16, torch.float16], ids=["bf16", "fp16"]) | ||
@pytest.mark.parametrize("qtype", [qint2, qint4], ids=["qint2", "qint4"]) | ||
@pytest.mark.parametrize("axis", [0, -1], ids=["first-axis", "last-axis"]) | ||
@pytest.mark.parametrize("group_size", [32, 64, 128]) | ||
def test_hqq_optimizer(input_shape, dtype, qtype, axis, group_size, device): | ||
a = random_tensor(input_shape, dtype=dtype).to(device) | ||
max_scale, max_shift = MaxOptimizer()(a, bits=qtype.bits, axis=axis, group_size=group_size) | ||
max_mean_error, max_sim = compare_quantized_tensor(a, qtype, axis, group_size, max_scale, max_shift) | ||
hqq_scale, hqq_shift = HqqOptimizer(verbose=True)(a, bits=qtype.bits, axis=axis, group_size=group_size) | ||
hqq_mean_error, hqq_sim = compare_quantized_tensor(a, qtype, axis, group_size, hqq_scale, hqq_shift) | ||
# HQQ optimizes the mean error, so it should be lower | ||
assert hqq_mean_error <= max_mean_error | ||
# HQQ cosine similarity should be also closer to 1 | ||
print(max_sim, hqq_sim) | ||
# assert torch.abs(1 - hqq_sim) <= torch.abs(1 - max_sim) |