Skip to content

Commit

Permalink
AQLM quantizer support (#28928)
Browse files Browse the repository at this point in the history
* aqlm init

* calibration and dtypes

* docs

* Readme update

* is_aqlm_available

* Simpler link in docs

* Test TODO real reference

* init _import_structure fix

* AqlmConfig autodoc

* integration aqlm

* integrations in tests

* docstring fix

* legacy typing

* Less typings

* More kernels information

* Performance -> Accuracy

* correct tests

* remoced multi-gpu test

* Update docs/source/en/quantization.md

Co-authored-by: Younes Belkada <[email protected]>

* Update src/transformers/utils/quantization_config.py

Co-authored-by: Arthur <[email protected]>

* Brought back multi-gpu tests

* Update src/transformers/integrations/aqlm.py

Co-authored-by: Marc Sun <[email protected]>

* Update tests/quantization/aqlm_integration/test_aqlm.py

Co-authored-by: Marc Sun <[email protected]>

---------

Co-authored-by: Andrei Panferov <[email protected]>
Co-authored-by: Younes Belkada <[email protected]>
Co-authored-by: Arthur <[email protected]>
Co-authored-by: Marc Sun <[email protected]>
  • Loading branch information
5 people authored Feb 14, 2024
1 parent 63ffd56 commit 1ecf5f7
Show file tree
Hide file tree
Showing 14 changed files with 489 additions and 2 deletions.
3 changes: 3 additions & 0 deletions docker/transformers-all-latest-gpu/Dockerfile
Original file line number Diff line number Diff line change
Expand Up @@ -55,6 +55,9 @@ RUN python3 -m pip install --no-cache-dir auto-gptq --extra-index-url https://hu
# Add einops for additional model testing
RUN python3 -m pip install --no-cache-dir einops

# Add aqlm for quantization testing
RUN python3 -m pip install --no-cache-dir aqlm[gpu]==1.0.1

# Add autoawq for quantization testing
RUN python3 -m pip install --no-cache-dir https://github.com/casper-hansen/AutoAWQ/releases/download/v0.1.8/autoawq-0.1.8+cu118-cp38-cp38-linux_x86_64.whl

Expand Down
4 changes: 4 additions & 0 deletions docs/source/en/main_classes/quantization.md
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,10 @@ Learn how to quantize models in the [Quantization](../quantization) guide.

</Tip>

## AqlmConfig

[[autodoc]] AqlmConfig

## AwqConfig

[[autodoc]] AwqConfig
Expand Down
28 changes: 28 additions & 0 deletions docs/source/en/quantization.md
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,34 @@ Interested in adding a new quantization method to Transformers? Read the [HfQuan

</Tip>

## AQLM



Try AQLM on [Google Colab](https://colab.research.google.com/drive/1-xZmBRXT5Fm3Ghn4Mwa2KRypORXb855X?usp=sharing)!

Additive Quantization of Language Models ([AQLM](https://arxiv.org/abs/2401.06118)) is a Large Language Models compression method. It quantizes multiple weights together and take advantage of interdependencies between them. AQLM represents groups of 8-16 weights as a sum of multiple vector codes.

Inference support for AQLM is realised in the `aqlm` library. Make sure to install it to run the models (note aqlm works only with python>=3.10):
```bash
pip install aqlm[gpu,cpu]
```

The library provides efficient kernels for both GPU and CPU inference.

The instructions on how to quantize models yourself, as well as all the relevant code can be found in the corresponding GitHub [repository](https://github.com/Vahe1994/AQLM).

### AQLM configurations

AQLM quantization setpus vary mainly on the number of codebooks used as well as codebook sizes in bits. The most popular setups, as well as inference kernels they support are:

| Kernel | Number of codebooks | Codebook size, bits | Notation | Accuracy | Speedup | Fast GPU inference | Fast CPU inference |
|---|---------------------|---------------------|----------|-------------|-------------|--------------------|--------------------|
| Triton | K | N | KxN | - | Up to ~0.7x |||
| CUDA | 1 | 16 | 1x16 | Best | Up to ~1.3x |||
| CUDA | 2 | 8 | 2x8 | OK | Up to ~3.0x |||
| Numba | K | 8 | Kx8 | Good | Up to ~4.0x |||

## AWQ

<Tip>
Expand Down
4 changes: 2 additions & 2 deletions src/transformers/__init__.py
Original file line number Diff line number Diff line change
Expand Up @@ -1087,7 +1087,7 @@
"is_vision_available",
"logging",
],
"utils.quantization_config": ["AwqConfig", "BitsAndBytesConfig", "GPTQConfig"],
"utils.quantization_config": ["AqlmConfig", "AwqConfig", "BitsAndBytesConfig", "GPTQConfig"],
}

# sentencepiece-backed objects
Expand Down Expand Up @@ -5845,7 +5845,7 @@
)

# bitsandbytes config
from .utils.quantization_config import AwqConfig, BitsAndBytesConfig, GPTQConfig
from .utils.quantization_config import AqlmConfig, AwqConfig, BitsAndBytesConfig, GPTQConfig

try:
if not is_sentencepiece_available():
Expand Down
2 changes: 2 additions & 0 deletions src/transformers/integrations/__init__.py
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,7 @@


_import_structure = {
"aqlm": ["replace_with_aqlm_linear"],
"awq": ["fuse_awq_modules", "replace_with_awq_linear"],
"bitsandbytes": [
"get_keys_to_not_convert",
Expand Down Expand Up @@ -80,6 +81,7 @@
}

if TYPE_CHECKING:
from .aqlm import replace_with_aqlm_linear
from .awq import fuse_awq_modules, replace_with_awq_linear
from .bitsandbytes import (
get_keys_to_not_convert,
Expand Down
99 changes: 99 additions & 0 deletions src/transformers/integrations/aqlm.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,99 @@
# 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.
"AQLM (Additive Quantization of Language Model) integration file"


from ..utils import is_accelerate_available, is_aqlm_available, is_torch_available


if is_torch_available():
import torch.nn as nn


def replace_with_aqlm_linear(
model,
quantization_config=None,
linear_weights_not_to_quantize=None,
current_key_name=None,
has_been_replaced=False,
):
"""
Public method that recursively replaces the Linear layers of the given model with AQLM quantized layers.
`accelerate` is needed to use this method. Returns the converted model and a boolean that indicates if the
conversion has been successfull or not.
Args:
model (`torch.nn.Module`):
The model to convert, can be any `torch.nn.Module` instance.
quantization_config (`AqlmConfig`):
The quantization config object that contains the quantization parameters.
linear_weights_not_to_quantize (`list[str]`, *optional*):
A list of nn.Linear weights to not convert. If a parameter path is in the list (e.g. `lm_head.weight`), the corresponding module will not be
converted.
current_key_name (`list`, *optional*):
A list that contains the current key name. This is used for recursion and should not be passed by the user.
has_been_replaced (`bool`, *optional*):
A boolean that indicates if the conversion has been successful or not. This is used for recursion and
should not be passed by the user.
"""
if not is_aqlm_available():
raise ValueError("AQLM is not available. Please install it with `pip install aqlm[cpu,gpu]`")

if not is_accelerate_available():
raise ValueError("AQLM requires Accelerate to be installed: `pip install accelerate`")

if linear_weights_not_to_quantize is None:
linear_weights_not_to_quantize = []

from accelerate import init_empty_weights
from aqlm import QuantizedLinear

for name, module in model.named_children():
if current_key_name is None:
current_key_name = []
current_key_name.append(name)

if isinstance(module, nn.Linear):
# Check if the current key is not in the `linear_weights_not_to_quantize`
if ".".join(current_key_name) + ".weight" not in linear_weights_not_to_quantize:
with init_empty_weights():
in_features = module.in_features
out_features = module.out_features

model._modules[name] = QuantizedLinear(
in_features,
out_features,
bias=module.bias is not None,
in_group_size=quantization_config.in_group_size,
out_group_size=quantization_config.out_group_size,
num_codebooks=quantization_config.num_codebooks,
nbits_per_codebook=quantization_config.nbits_per_codebook,
)
has_been_replaced = True

# Store the module class in case we need to transpose the weight later
model._modules[name].source_cls = type(module)
# Force requires grad to False to avoid unexpected errors
model._modules[name].requires_grad_(False)
if len(list(module.children())) > 0:
_, has_been_replaced = replace_with_aqlm_linear(
module,
quantization_config=quantization_config,
linear_weights_not_to_quantize=linear_weights_not_to_quantize,
current_key_name=current_key_name,
has_been_replaced=has_been_replaced,
)
# Remove the last key for recursion
current_key_name.pop(-1)
return model, has_been_replaced
4 changes: 4 additions & 0 deletions src/transformers/quantizers/auto.py
Original file line number Diff line number Diff line change
Expand Up @@ -16,12 +16,14 @@

from ..models.auto.configuration_auto import AutoConfig
from ..utils.quantization_config import (
AqlmConfig,
AwqConfig,
BitsAndBytesConfig,
GPTQConfig,
QuantizationConfigMixin,
QuantizationMethod,
)
from .quantizer_aqlm import AqlmHfQuantizer
from .quantizer_awq import AwqQuantizer
from .quantizer_bnb_4bit import Bnb4BitHfQuantizer
from .quantizer_bnb_8bit import Bnb8BitHfQuantizer
Expand All @@ -33,13 +35,15 @@
"bitsandbytes_4bit": Bnb4BitHfQuantizer,
"bitsandbytes_8bit": Bnb8BitHfQuantizer,
"gptq": GptqHfQuantizer,
"aqlm": AqlmHfQuantizer,
}

AUTO_QUANTIZATION_CONFIG_MAPPING = {
"awq": AwqConfig,
"bitsandbytes_4bit": BitsAndBytesConfig,
"bitsandbytes_8bit": BitsAndBytesConfig,
"gptq": GPTQConfig,
"aqlm": AqlmConfig,
}


Expand Down
89 changes: 89 additions & 0 deletions src/transformers/quantizers/quantizer_aqlm.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,89 @@
# 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 ..integrations import replace_with_aqlm_linear
from ..utils import is_accelerate_available, is_aqlm_available, is_torch_available, logging
from ..utils.quantization_config import QuantizationConfigMixin


if is_torch_available():
import torch

logger = logging.get_logger(__name__)


class AqlmHfQuantizer(HfQuantizer):
"""
Quantizer of the AQLM method. Enables the loading of prequantized models.
"""

requires_calibration = True
required_packages = ["aqlm"]
optimum_quantizer = None

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 `aqlm` quantization requires Accelerate: `pip install accelerate`")

if not is_aqlm_available():
raise ImportError("Using `aqlm` quantization requires AQLM: `pip install aqlm[gpu,cpu]`")

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 AQLM inference on GPU and loading the model in `torch.float16`. To overwrite it, set `torch_dtype` manually."
)
else:
torch_dtype = torch.float32
logger.info(
"CUDA is unavailable. Assuming AQLM inference on CPU and loading the model in `torch.float32`. To overwrite it, set `torch_dtype` manually."
)
return torch_dtype

def _process_model_before_weight_loading(
self,
model: "PreTrainedModel",
**kwargs,
):
replace_with_aqlm_linear(
model,
quantization_config=self.quantization_config,
linear_weights_not_to_quantize=self.quantization_config.linear_weights_not_to_quantize,
)
model.config.quantization_config = self.quantization_config

def _process_model_after_weight_loading(self, model: "PreTrainedModel", **kwargs):
model._is_quantized_training_enabled = False
return model

@property
def is_trainable(self, model: Optional["PreTrainedModel"] = None):
return False

@property
def is_serializable(self):
return True
8 changes: 8 additions & 0 deletions src/transformers/testing_utils.py
Original file line number Diff line number Diff line change
Expand Up @@ -53,6 +53,7 @@
from .utils import (
is_accelerate_available,
is_apex_available,
is_aqlm_available,
is_auto_awq_available,
is_auto_gptq_available,
is_bitsandbytes_available,
Expand Down Expand Up @@ -956,6 +957,13 @@ def require_apex(test_case):
return unittest.skipUnless(is_apex_available(), "test requires apex")(test_case)


def require_aqlm(test_case):
"""
Decorator marking a test that requires aqlm
"""
return unittest.skipUnless(is_aqlm_available(), "test requires aqlm")(test_case)


def require_bitsandbytes(test_case):
"""
Decorator for bits and bytes (bnb) dependency
Expand Down
1 change: 1 addition & 0 deletions src/transformers/utils/__init__.py
Original file line number Diff line number Diff line change
Expand Up @@ -105,6 +105,7 @@
get_torch_version,
is_accelerate_available,
is_apex_available,
is_aqlm_available,
is_auto_awq_available,
is_auto_gptq_available,
is_bitsandbytes_available,
Expand Down
5 changes: 5 additions & 0 deletions src/transformers/utils/import_utils.py
Original file line number Diff line number Diff line change
Expand Up @@ -74,6 +74,7 @@ def _is_package_available(pkg_name: str, return_version: bool = False) -> Union[

_accelerate_available, _accelerate_version = _is_package_available("accelerate", return_version=True)
_apex_available = _is_package_available("apex")
_aqlm_available = _is_package_available("aqlm")
_bitsandbytes_available = _is_package_available("bitsandbytes")
# `importlib.metadata.version` doesn't work with `bs4` but `beautifulsoup4`. For `importlib.util.find_spec`, reversed.
_bs4_available = importlib.util.find_spec("bs4") is not None
Expand Down Expand Up @@ -570,6 +571,10 @@ def is_apex_available():
return _apex_available


def is_aqlm_available():
return _aqlm_available


def is_ninja_available():
r"""
Code comes from *torch.utils.cpp_extension.is_ninja_available()*. Returns `True` if the
Expand Down
Loading

0 comments on commit 1ecf5f7

Please sign in to comment.