Repository: yuhuixu1993/qa-lora
Branch: main
Commit: 91604c71e981
Files: 7
Total size: 58.3 KB
Directory structure:
gitextract_jckgmwf6/
├── LICENSE
├── README.md
├── environment.yaml
├── merge.py
├── peft_utils.py
├── qalora.py
└── requirements.txt
================================================
FILE CONTENTS
================================================
================================================
FILE: LICENSE
================================================
MIT License
Copyright (c) 2023 YuhuiXu
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
================================================
FILE: README.md
================================================
# QA-LoRA
QA-LoRA has been accepted by ICLR 2024!
This repository provides the official PyTorch implementation of [QA-LoRA: Quantization-Aware Low-Rank Adaptation of Large Language Models](https://arxiv.org/pdf/2309.14717.pdf).
<div align="center">
<img src="image/qalora.png" width="600"/>
</div>
QA-LoRA is easily implemented with a few lines of code, and it equips the original LoRA with two-fold abilities: (i) during fine-tuning, the LLM's weights are quantized (e.g., into INT4) to reduce time and memory usage; (ii) after fine-tuning, the LLM and auxiliary weights are naturally integrated into a quantized model without loss of accuracy.
## Todo list
Fix the conflict with the newest Auto-gptq version.
## Installation
```bash
conda create -n qalora python=3.8
conda activate qalora
conda install pytorch==2.2.0 torchvision==0.17.0 torchaudio==2.2.0 pytorch-cuda=12.1 -c pytorch -c nvidia
git clone -b v0.3.0 https://github.com/PanQiWei/AutoGPTQ.git && cd AutoGPTQ
pip install .
cd ..
pip install bitsandbytes
pip install -r requirements.txt
pip install protobuf==3.20.*
```
Change the `peft_utils.py` in your own auto-gptq path(python path/auto_gptq/utils/peft_utils.py) with the new one.
For the users of [GPTQLORA](https://github.com/qwopqwop200/gptqlora), you only need to change the `peft_utils.py` file.
## Quantization
We use [GPTQ](https://github.com/qwopqwop200/GPTQ-for-LLaMa) for quantization.
bits=4, group-size=32, act-order=False
If you change the group-size, you need to change the group_size in `peft_utils.py` and `merge.py` accordingly.
## Training
```bash
python qalora.py --model_path <path>
```
The file structure of the model checkpoint is as follows:
```
config.json llama7b-4bit-32g.bin special_tokens_map.json tokenizer_config.json
generation_config.json quantize_config.json tokenizer.model
```
## Merge
Note that our trained LoRA modules can be perfectly merged into the quantized model. We offer a simple merged script in this repo.
## Notice
### About the implementations
There are two kinds of implementations of the dimention reduction(x from D_in to D_in//L). Both are mathematical equivalent.
#### The first one(this repo)
Adopt avgpooling operation. But the weights of adapters will be divided by D_in//L during merge(refer to `merge.py`).
```bash
adapter_result = (lora_B(lora_A(lora_dropout(self.qa_pool(x)))） * scale).type_as(result)
model[tmp_key+'.qzeros'] -= (lora['base_model.model.'+tmp_key+'.lora_B.weight'] @ lora['base_model.model.'+tmp_key+'.lora_A.weight']).t() * scale / group_size / model[tmp_key+'.scales']
```
#### The second one
Utilize sum operation. The adapters do not need to be divided during merge)
```bash
adapter_result = (lora_B(lora_A(lora_dropout(self.qa_pool(x) * group_size))） * scale).type_as(result)
model[tmp_key+'.qzeros'] -= (lora['base_model.model.'+tmp_key+'.lora_B.weight'] @ lora['base_model.model.'+tmp_key+'.lora_A.weight']).t() * scale / model[tmp_key+'.scales']
```
### About the quantization
Some GPTQ implementation such as [GPTQ-for-llama](https://github.com/qwopqwop200/GPTQ-for-LLaMa) further compress the zeros into qzeros. You need to decode the qzeros first and restore fp16 format zeros.
## Acknowledgements
Our code is based on [QLoRA](https://github.com/artidoro/qlora), [GPTQLORA](https://github.com/qwopqwop200/gptqlora), [Auto-GPTQ](https://github.com/PanQiWei/AutoGPTQ/tree/main)
================================================
FILE: environment.yaml
================================================
name: alpaca
channels:
- defaults
dependencies:
- _libgcc_mutex=0.1=main
- _openmp_mutex=5.1=1_gnu
- ca-certificates=2023.05.30=h06a4308_0
- ld_impl_linux-64=2.38=h1181459_1
- libffi=3.4.4=h6a678d5_0
- libgcc-ng=11.2.0=h1234567_1
- libgomp=11.2.0=h1234567_1
- libstdcxx-ng=11.2.0=h1234567_1
- ncurses=6.4=h6a678d5_0
- openssl=3.0.9=h7f8727e_0
- pip=23.1.2=py38h06a4308_0
- python=3.8.16=h955ad1f_4
- readline=8.2=h5eee18b_0
- setuptools=67.8.0=py38h06a4308_0
- sqlite=3.41.2=h5eee18b_0
- tk=8.6.12=h1ccaba5_0
- wheel=0.38.4=py38h06a4308_0
- xz=5.4.2=h5eee18b_0
- zlib=1.2.13=h5eee18b_0
- pip:
- absl-py==1.2.0
- accelerate==0.21.0.dev0
- addict==2.4.0
- aiohttp==3.8.4
- aiosignal==1.3.1
- appdirs==1.4.4
- async-timeout==4.0.2
- attrs==22.2.0
- auto-gptq==0.3.0.dev0
- bert-score==0.3.13
- bitsandbytes==0.39.0
- certifi==2022.9.24
- charset-normalizer==3.0.1
- click==8.1.3
- cmake==3.26.3
- colorama==0.4.5
- contourpy==1.0.6
- cycler==0.11.0
- datasets==2.12.0
- dill==0.3.5.1
- evaluate==0.4.0
- fonttools==4.39.4
- frozenlist==1.3.3
- fsspec==2022.11.0
- gitdb==4.0.10
- gitpython==3.1.31
- huggingface-hub==0.14.1
- importlib-resources==5.12.0
- jinja2==3.1.2
- joblib==1.2.0
- lazy-import==0.2.2
- lit==16.0.5
- lxml==4.8.0
- markupsafe==2.1.2
- matplotlib==3.7.1
- mpmath==1.2.1
- multidict==6.0.2
- multiprocess==0.70.12.2
- networkx==2.8.8
- ninja==1.11.1
- nltk==3.8.1
- numpy==1.24.2
- packaging==23.1
- pandas==2.0.0
- pathlib2==2.3.7.post1
- pathtools==0.1.2
- peft==0.4.0.dev0
- pillow==9.3.0
- protobuf==3.20.2
- psutil==5.9.4
- pyarrow==10.0.1
- pyparsing==3.0.9
- python-dateutil==2.8.2
- pytz==2022.6
- pyyaml==6.0
- requests==2.31.0
- responses==0.18.0
- rouge==1.0.0
- rouge-score==0.1.2
- safetensors==0.3.1
- scikit-learn==1.2.2
- scipy==1.10.1
- sentencepiece==0.1.99
- sentry-sdk==1.24.0
- setproctitle==1.3.2
- six==1.16.0
- smmap==5.0.0
- sympy==1.11.1
- terminaltables==3.1.10
- threadpoolctl==3.1.0
- tokenizers==0.13.3
- torch==2.0.0+cu117
- tqdm==4.63.1
- transformers==4.31.0.dev0
- triton==2.0.0
- typing-extensions==4.6.2
- tzdata==2022.7
- urllib3==1.26.16
- wandb==0.15.2
- xformers==0.0.20
- xxhash==3.0.0
- yapf==0.32.0
- yarl==1.8.1
- zipp==3.15.0
================================================
FILE: merge.py
================================================
import torch
model_path = 'path of the quantized model'
lora_path = 'path of the saved LoRA adapters'
merged_path = 'target path of the merged model'
scale = 16 /64
group_size = 32
model = torch.load(model_path, map_location='cpu')
lora = torch.load(lora_path, map_location='cpu')
tmp_keys = [key[17:-14] for key in lora.keys() if 'lora_A' in key]
for tmp_key in tmp_keys:
model[tmp_key+'.qzeros'] -= (lora['base_model.model.'+tmp_key+'.lora_B.weight'] @ lora['base_model.model.'+tmp_key+'.lora_A.weight']).t() * scale / group_size /model[tmp_key+'.scales']
torch.save(model, merged_path)
================================================
FILE: peft_utils.py
================================================
import warnings
import re
from contextlib import contextmanager
from dataclasses import asdict
from enum import Enum
from typing import List, Optional
import torch
from peft import get_peft_model, PeftConfig, PeftModel, PeftType
from peft.peft_model import PEFT_TYPE_TO_MODEL_MAPPING
from peft.tuners.lora import LoraConfig, LoraLayer, LoraModel, Embedding
from peft.tuners.adalora import AdaLoraConfig, AdaLoraLayer, AdaLoraModel
from peft.mapping import PEFT_TYPE_TO_CONFIG_MAPPING
from peft.utils.other import _get_submodules
from ..modeling._base import BaseGPTQForCausalLM
group_size = 32 # quantization group_size
class GPTQLoraConfig(LoraConfig):
injected_fused_attention: bool = False
injected_fused_mlp: bool = False
class GPTQLoraLinear(torch.nn.Linear, LoraLayer):
def __init__(
self,
adapter_name: str,
linear_module: torch.nn.Linear,
r: int = 0,
lora_alpha: int = 1,
lora_dropout: float = 0.0,
fan_in_fan_out: bool = False, # Set this to True if the layer to replace stores weight like (fan_in, fan_out)
**kwargs,
):
init_lora_weights = kwargs.pop("init_lora_weights", True)
torch.nn.Linear.__init__(self, linear_module.in_features, linear_module.out_features)
LoraLayer.__init__(self, linear_module.in_features//group_size, linear_module.out_features)
self.linear_module = linear_module
self.weight.requires_grad = False
self.weight = self.linear_module.weight
self.bias = self.linear_module.bias
self.fan_in_fan_out = fan_in_fan_out
if fan_in_fan_out:
self.weight.data = self.weight.data.T
self.update_layer(adapter_name, r, lora_alpha, lora_dropout, init_lora_weights)
self.active_adapter = adapter_name
self.qa_pool = torch.nn.AvgPool1d(group_size) # using pooling layer to conduct sum operation
def reset_lora_parameters(self, adapter_name):
if adapter_name in self.lora_A.keys():
torch.nn.init.xavier_uniform_(self.lora_A[adapter_name].weight)
torch.nn.init.zeros_(self.lora_B[adapter_name].weight)
def merge(self):
raise NotImplementedError("gptq model not support merge lora adapter")
def unmerge(self):
raise NotImplementedError("gptq model not support unmerge lora adapter")
def forward(self, x: torch.Tensor):
previous_dtype = x.dtype
if self.active_adapter not in self.lora_A.keys():
return self.linear_module(x)
if self.disable_adapters:
if self.r[self.active_adapter] > 0 and self.merged:
self.unmerge()
result = self.linear_module(x)
elif self.r[self.active_adapter] > 0 and not self.merged:
result = self.linear_module(x)
lora_B = self.lora_B[self.active_adapter]
lora_A = self.lora_A[self.active_adapter]
lora_dropout = self.lora_dropout[self.active_adapter]
scale = self.scaling[self.active_adapter]
x = x.type_as(lora_A.weight.data)
adapter_result = (lora_B(lora_A(lora_dropout(self.qa_pool(x)))) * scale).type_as(result)
result += adapter_result
else:
result = self.linear_module(x)
result = result.to(previous_dtype)
return result
class GPTQLoraModel(LoraModel):
def _find_and_replace(self, adapter_name):
lora_config = self.peft_config[adapter_name]
is_target_modules_in_base_model = False
kwargs = {
"r": lora_config.r,
"lora_alpha": lora_config.lora_alpha,
"lora_dropout": lora_config.lora_dropout,
"fan_in_fan_out": lora_config.fan_in_fan_out,
"init_lora_weights": lora_config.init_lora_weights,
}
key_list = [key for key, _ in self.model.named_modules()]
for key in key_list:
if isinstance(lora_config.target_modules, str):
target_module_found = re.fullmatch(lora_config.target_modules, key)
else:
target_module_found = any(key.endswith(target_key) for target_key in lora_config.target_modules)
if target_module_found:
if not is_target_modules_in_base_model:
is_target_modules_in_base_model = True
parent, target, target_name = _get_submodules(self.model, key)
bias = False
if hasattr(target, "bias"):
bias = target.bias is not None
if isinstance(target, LoraLayer):
target.update_layer(
adapter_name,
lora_config.r,
lora_config.lora_alpha,
lora_config.lora_dropout,
lora_config.init_lora_weights,
)
else:
if isinstance(target, torch.nn.Embedding):
embedding_kwargs = kwargs.copy()
embedding_kwargs.pop("fan_in_fan_out", None)
in_features, out_features = target.num_embeddings, target.embedding_dim
new_module = Embedding(adapter_name, in_features, out_features, **embedding_kwargs)
else:
if isinstance(target, torch.nn.Linear):
if kwargs["fan_in_fan_out"]:
warnings.warn(
"fan_in_fan_out is set to True but the target module is `torch.nn.Linear`. "
"Setting fan_in_fan_out to False."
)
kwargs["fan_in_fan_out"] = lora_config.fan_in_fan_out = False
else:
raise ValueError(
f"Target module {target} is not supported. "
f"Currently, only `torch.nn.Linear` and its subclasses are supported."
)
new_module = GPTQLoraLinear(adapter_name, target, **kwargs)
self._replace_module(parent, target_name, new_module, target)
if not is_target_modules_in_base_model:
raise ValueError(
f"Target modules {lora_config.target_modules} not found in the base model. "
f"Please check the target modules and try again."
)
def _replace_module(self, parent_module, child_name, new_module, old_module):
setattr(parent_module, child_name, new_module)
if not isinstance(new_module, GPTQLoraLinear):
new_module.weight = old_module.weight
if hasattr(old_module, "bias"):
if old_module.bias is not None:
new_module.bias = old_module.bias
if getattr(old_module, "state", None) is not None:
new_module.state = old_module.state
new_module.to(old_module.weight.device)
# dispatch to correct device
for name, module in new_module.named_modules():
if "lora_" in name:
module.to(old_module.weight.device)
def merge_adapter(self):
raise NotImplementedError("gptq model not support merge ada lora adapter")
def unmerge_adapter(self):
raise NotImplementedError("gptq model not support unmerge ada lora adapter")
def merge_and_unload(self):
raise NotImplementedError("gptq model not support merge and unload")
class GPTQAdaLoraConfig(AdaLoraConfig):
injected_fused_attention: bool = False
injected_fused_mlp: bool = False
class GPTQSVDLinear(torch.nn.Linear, AdaLoraLayer):
def __init__(
self,
adapter_name: str,
linear_module: torch.nn.Linear,
r: int = 0,
lora_alpha: int = 1,
lora_dropout: float = 0.0,
fan_in_fan_out: bool = False, # Set this to True if the layer to replace stores weight like (fan_in, fan_out)
**kwargs,
):
init_lora_weights = kwargs.pop("init_lora_weights", True)
torch.nn.Linear.__init__(self, linear_module.in_features, linear_module.out_features)
AdaLoraLayer.__init__(self, linear_module.in_features, linear_module.out_features)
self.linear_module = linear_module
self.weight.requires_grad = False
self.weight = self.linear_module.weight
self.bias = self.linear_module.bias
self.fan_in_fan_out = fan_in_fan_out
if fan_in_fan_out:
self.weight.data = self.weight.data.T
self.update_layer(adapter_name, r, lora_alpha, lora_dropout, init_lora_weights)
self.active_adapter = adapter_name
def merge(self):
raise NotImplementedError("gptq model not support merge lora adapter")
def unmerge(self):
raise NotImplementedError("gptq model not support unmerge lora adapter")
def forward(self, x: torch.Tensor):
if self.active_adapter not in self.lora_A.keys():
return self.linear_module(x)
if self.disable_adapters:
if self.r[self.active_adapter] > 0 and self.merged:
self.unmerge()
result = self.linear_module(x)
elif self.r[self.active_adapter] > 0 and not self.merged:
result = self.linear_module(x)
result += (
(
self.lora_dropout[self.active_adapter](x)
@ (self.lora_A[self.active_adapter] * self.lora_E[self.active_adapter]).T
@ self.lora_B[self.active_adapter].T
)
* self.scaling[self.active_adapter]
/ (self.ranknum[self.active_adapter] + 1e-5)
)
else:
result = self.linear_module(x)
return result
class GPTQAdaLoraModel(AdaLoraModel):
def _find_and_replace(self, adapter_name):
lora_config = self.peft_config[adapter_name]
is_target_modules_in_base_model = False
kwargs = {
"r": lora_config.init_r,
"lora_alpha": lora_config.lora_alpha,
"lora_dropout": lora_config.lora_dropout,
"fan_in_fan_out": lora_config.fan_in_fan_out,
"init_lora_weights": lora_config.init_lora_weights,
}
key_list = [key for key, _ in self.model.named_modules()]
for key in key_list:
if isinstance(lora_config.target_modules, str):
target_module_found = re.fullmatch(lora_config.target_modules, key)
else:
target_module_found = any(key.endswith(target_key) for target_key in lora_config.target_modules)
if target_module_found:
if not is_target_modules_in_base_model:
is_target_modules_in_base_model = True
parent, target, target_name = _get_submodules(self.model, key)
bias = target.bias is not None
if isinstance(target, LoraLayer):
target.update_layer(
adapter_name,
lora_config.init_r,
lora_config.lora_alpha,
lora_config.lora_dropout,
lora_config.init_lora_weights,
)
else:
if isinstance(target, torch.nn.Linear):
in_features, out_features = target.in_features, target.out_features
if kwargs["fan_in_fan_out"]:
warnings.warn(
"fan_in_fan_out is set to True but the target module is `torch.nn.Linear`. "
"Setting fan_in_fan_out to False."
)
kwargs["fan_in_fan_out"] = lora_config.fan_in_fan_out = False
else:
raise ValueError(
f"Target module {target} is not supported. "
f"Currently, only `torch.nn.Linear` and its subclasses are supported."
)
new_module = GPTQSVDLinear(adapter_name, target, **kwargs)
self._replace_module(parent, target_name, new_module, target)
if not is_target_modules_in_base_model:
raise ValueError(
f"Target modules {lora_config.target_modules} not found in the base model. "
f"Please check the target modules and try again."
)
def _replace_module(self, parent_module, child_name, new_module, old_module):
setattr(parent_module, child_name, new_module)
# dispatch to correct device
for name, module in new_module.named_modules():
if "lora_" in name:
module.to(old_module.weight.device)
def merge_adapter(self):
raise NotImplementedError("gptq model not support merge ada lora adapter")
def unmerge_adapter(self):
raise NotImplementedError("gptq model not support unmerge ada lora adapter")
def merge_and_unload(self):
raise NotImplementedError("gptq model not support merge and unload")
def find_all_linear_names(model: BaseGPTQForCausalLM, ignore: Optional[List[str]] = None, ignore_lm_head: bool = True):
if not ignore:
ignore = []
lm_head_name = model.lm_head_name
if ignore_lm_head and lm_head_name not in ignore:
ignore.append(lm_head_name)
results = set()
for n, m in model.named_modules():
if isinstance(m, torch.nn.Linear):
res = n.split('.')[-1]
if res not in ignore:
results.add(res)
return list(results)
@contextmanager
def hijack_peft_mappings():
PEFT_TYPE_TO_CONFIG_MAPPING[PeftType.LORA] = GPTQLoraConfig
PEFT_TYPE_TO_MODEL_MAPPING[PeftType.LORA] = GPTQLoraModel
PEFT_TYPE_TO_CONFIG_MAPPING[PeftType.ADALORA] = GPTQAdaLoraConfig
PEFT_TYPE_TO_MODEL_MAPPING[PeftType.ADALORA] = GPTQAdaLoraModel
try:
yield
except:
PEFT_TYPE_TO_CONFIG_MAPPING[PeftType.LORA] = GPTQLoraConfig
PEFT_TYPE_TO_MODEL_MAPPING[PeftType.LORA] = GPTQLoraModel
PEFT_TYPE_TO_CONFIG_MAPPING[PeftType.ADALORA] = GPTQAdaLoraConfig
PEFT_TYPE_TO_MODEL_MAPPING[PeftType.ADALORA] = GPTQAdaLoraModel
raise
finally:
PEFT_TYPE_TO_CONFIG_MAPPING[PeftType.LORA] = GPTQLoraConfig
PEFT_TYPE_TO_MODEL_MAPPING[PeftType.LORA] = GPTQLoraModel
PEFT_TYPE_TO_CONFIG_MAPPING[PeftType.ADALORA] = GPTQAdaLoraConfig
PEFT_TYPE_TO_MODEL_MAPPING[PeftType.ADALORA] = GPTQAdaLoraModel
def get_gptq_peft_model(
model: BaseGPTQForCausalLM,
peft_config: PeftConfig = None,
model_id: str = None,
adapter_name: str = "default",
auto_find_all_linears: bool = True,
train_mode: bool = False
):
if train_mode and not model.trainable:
model.enable_trainable_mode()
if train_mode and not peft_config:
raise ValueError("peft_config not specified when in train mode.")
if not train_mode and not model_id:
raise ValueError("model_id(where to load adapters) not specified when in inference mode.")
if model.fused_attn_module_type is not None and not model.injected_fused_attention:
peft_types = [PeftType.LORA.value, PeftType.ADALORA.value]
warnings.warn(
f"You can just ignore this warning if the peft type you use isn't in {peft_types}.\n"
f"{model.__class__.__name__} supports injecting fused attention but not enables this time. "
"If you are training adapters, you must also disable fused attention injection when loading quantized "
"base model at inference time, otherwise adapters may not be added to base model properly. "
"If you are loading adapters to do inference, you can reference to adapter's config file to check "
"whether the adapters are trained using base model that not enable fused attention injection."
)
if model.injected_fused_mlp:
raise NotImplementedError("GPTQ model that enables fused mlp injection is not supported to integrate with peft.")
if train_mode:
peft_type = peft_config.peft_type
if not isinstance(peft_type, str):
peft_type = peft_type.value
if peft_type in [PeftType.LORA.value, PeftType.ADALORA.value]:
if auto_find_all_linears:
peft_config.target_modules = find_all_linear_names(model, ignore_lm_head=True)
if peft_type == PeftType.LORA.value and not isinstance(peft_config, GPTQLoraConfig):
peft_config = GPTQLoraConfig(**peft_config.to_dict())
if peft_type == PeftType.ADALORA.value and not isinstance(peft_config, GPTQAdaLoraConfig):
peft_config = GPTQAdaLoraConfig(**peft_config.to_dict())
peft_config.injected_fused_attention = model.injected_fused_attention
peft_config.injected_fused_mlp = model.injected_fused_mlp
if peft_type == PeftType.ADAPTION_PROMPT.value:
if peft_config.adapter_layers > model.config.num_hidden_layers:
warnings.warn(
f"model has only {model.config.num_hidden_layers} layers "
f"but adapter_layers is set to {peft_config.adapter_layers}, "
f"will reset value to {model.config.num_hidden_layers}."
)
peft_config.adapter_layers = model.config.num_hidden_layers
if model.injected_fused_attention:
raise NotImplementedError(
"model with fused attention injected isn't supported to use ADAPTION_PROMPT peft type yet."
)
with hijack_peft_mappings():
try:
if train_mode:
peft_model = get_peft_model(model.model, peft_config, adapter_name=adapter_name)
else:
peft_model = PeftModel.from_pretrained(model.model, model_id, adapter_name)
except:
raise NotImplementedError(
f"{model.__class__.__name__} not support {peft_config.peft_type.value} peft type yet."
)
return peft_model
__all__ = [
"GPTQLoraConfig",
"GPTQLoraModel",
"GPTQAdaLoraConfig",
"GPTQAdaLoraModel",
"find_all_linear_names",
"get_gptq_peft_model"
]
================================================
FILE: qalora.py
================================================
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
from collections import defaultdict
import copy
import json
import os
from os.path import exists, join, isdir
from dataclasses import dataclass, field
import sys
from typing import Optional, Dict, Sequence
import numpy as np
from tqdm import tqdm
import logging
import torch
import transformers
from torch.nn.utils.rnn import pad_sequence
import argparse
from transformers import (
AutoTokenizer,
AutoModelForCausalLM,
set_seed,
Seq2SeqTrainer,
LlamaTokenizerFast
)
from datasets import load_dataset
import evaluate
from peft import (
LoraConfig,
get_peft_model_state_dict,
set_peft_model_state_dict,
PeftModel
)
from peft.tuners.lora import LoraLayer
from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR
from auto_gptq.utils.peft_utils import get_gptq_peft_model, GPTQLoraConfig
from auto_gptq import AutoGPTQForCausalLM
from auto_gptq.nn_modules.qlinear import GeneralQuantLinear
torch.backends.cuda.matmul.allow_tf32 = True
logger = logging.getLogger(__name__)
IGNORE_INDEX = -100
DEFAULT_PAD_TOKEN = "[PAD]"
import os
os.environ["TOKENIZERS_PARALLELISM"] = "false"
def prepare_model_for_int8_training(model, use_gradient_checkpointing=True):
r"""
This method wraps the entire protocol for preparing a model before running a training. This includes:
1- Cast the layernorm in fp32 2- making output embedding layer require grads 3- Add the upcasting of the lm
head to fp32
Args:
model, (`transformers.PreTrainedModel`):
The loaded model from `transformers`
"""
for name, param in model.named_parameters():
# freeze base model's layers
param.requires_grad = False
if use_gradient_checkpointing:
# For backward compatibility
if hasattr(model, "enable_input_require_grads"):
model.enable_input_require_grads()
else:
def make_inputs_require_grad(module, input, output):
output.requires_grad_(True)
model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)
# enable gradient checkpointing for memory efficiency
model.gradient_checkpointing_enable()
model.lm_head = model.lm_head.float()
for _, param in model.named_parameters():
if param.dtype == torch.float16:
param = param.float()
return model
@dataclass
class ModelArguments:
model_path: Optional[str] = field(
default="./llama-7b/"
)
trust_remote_code: Optional[bool] = field(
default=False,
metadata={"help": "Enable unpickling of arbitrary code in AutoModelForCausalLM#from_pretrained."}
)
@dataclass
class DataArguments:
eval_dataset_size: int = field(
default=1024, metadata={"help": "Size of validation dataset."}
)
max_train_samples: Optional[int] = field(
default=None,
metadata={
"help": "For debugging purposes or quicker training, truncate the number of training examples to this "
"value if set."
},
)
max_eval_samples: Optional[int] = field(
default=None,
metadata={
"help": "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
"value if set."
},
)
source_max_len: int = field(
default=1024,
metadata={"help": "Maximum source sequence length. Sequences will be right padded (and possibly truncated)."},
)
target_max_len: int = field(
default=256,
metadata={"help": "Maximum target sequence length. Sequences will be right padded (and possibly truncated)."},
)
dataset: str = field(
default='alpaca',
metadata={"help": "Which dataset to finetune on. See datamodule for options."}
)
@dataclass
class TrainingArguments(transformers.Seq2SeqTrainingArguments):
cache_dir: Optional[str] = field(
default=None
)
train_on_source: Optional[bool] = field(
default=False,
metadata={"help": "Whether to train on the input in addition to the target text."}
)
mmlu_split: Optional[str] = field(
default='eval',
metadata={"help": "The MMLU split to run on"}
)
mmlu_dataset: Optional[str] = field(
default='mmlu-fs',
metadata={"help": "MMLU dataset to use: options are `mmlu-zs` for zero-shot or `mmlu-fs` for few shot."}
)
do_mmlu_eval: Optional[bool] = field(
default=False,
metadata={"help": "Whether to run the MMLU evaluation."}
)
max_mmlu_samples: Optional[int] = field(
default=None,
metadata={"help": "If set, only evaluates on `max_mmlu_samples` of the MMMLU dataset."}
)
mmlu_source_max_len: int = field(
default=2048,
metadata={"help": "Maximum source sequence length for mmlu."}
)
full_finetune: bool = field(
default=False,
metadata={"help": "Finetune the entire model without adapters."}
)
adam8bit: bool = field(
default=False,
metadata={"help": "Use 8-bit adam."}
)
lora_r: int = field(
default=64,
metadata={"help": "Lora R dimension."}
)
lora_alpha: float = field(
default=16,
metadata={"help": " Lora alpha."}
)
lora_dropout: float = field(
default=0.0,
metadata={"help":"Lora dropout."}
)
max_memory_MB: int = field(
default=24000,
metadata={"help": "Free memory per gpu."}
)
report_to: str = field(
default='none',
metadata={"help": "To use wandb or something else for reporting."}
)
output_dir: str = field(default='./output', metadata={"help": 'The output dir for logs and checkpoints'})
optim: str = field(default='paged_adamw_32bit', metadata={"help": 'The optimizer to be used'})
per_device_train_batch_size: int = field(default=1, metadata={"help": 'The training batch size per GPU. Increase for better speed.'})
gradient_accumulation_steps: int = field(default=16, metadata={"help": 'How many gradients to accumulate before to perform an optimizer step'})
max_steps: int = field(default=10000, metadata={"help": 'How many optimizer update steps to take'})
weight_decay: float = field(default=0.0, metadata={"help": 'The L2 weight decay rate of AdamW'}) # use lora dropout instead for regularization if needed
learning_rate: float = field(default=0.0002, metadata={"help": 'The learnign rate'})
remove_unused_columns: bool = field(default=False, metadata={"help": 'Removed unused columns. Needed to make this codebase work.'})
max_grad_norm: float = field(default=0.3, metadata={"help": 'Gradient clipping max norm. This is tuned and works well for all models tested.'})
gradient_checkpointing: bool = field(default=True, metadata={"help": 'Use gradient checkpointing. You want to use this.'})
do_train: bool = field(default=True, metadata={"help": 'To train or not to train, that is the question?'})
lr_scheduler_type: str = field(default='constant', metadata={"help": 'Learning rate schedule. Constant a bit better than cosine, and has advantage for analysis'})
warmup_ratio: float = field(default=0.03, metadata={"help": 'Fraction of steps to do a warmup for'})
logging_steps: int = field(default=10, metadata={"help": 'The frequency of update steps after which to log the loss'})
group_by_length: bool = field(default=True, metadata={"help": 'Group sequences into batches with same length. Saves memory and speeds up training considerably.'})
save_strategy: str = field(default='steps', metadata={"help": 'When to save checkpoints'})
save_steps: int = field(default=250, metadata={"help": 'How often to save a model'})
save_total_limit: int = field(default=40, metadata={"help": 'How many checkpoints to save before the oldest is overwritten'})
@dataclass
class GenerationArguments:
# For more hyperparameters check:
# https://huggingface.co/docs/transformers/main_classes/text_generation#transformers.GenerationConfig
# Length arguments
max_new_tokens: Optional[int] = field(
default=256,
metadata={"help": "Maximum number of new tokens to be generated in evaluation or prediction loops"
"if predict_with_generate is set."}
)
min_new_tokens : Optional[int] = field(
default=None,
metadata={"help": "Minimum number of new tokens to generate."}
)
# Generation strategy
do_sample: Optional[bool] = field(default=False)
num_beams: Optional[int] = field(default=1)
num_beam_groups: Optional[int] = field(default=1)
penalty_alpha: Optional[float] = field(default=None)
use_cache: Optional[bool] = field(default=True)
# Hyperparameters for logit manipulation
temperature: Optional[float] = field(default=1.0)
top_k: Optional[int] = field(default=50)
top_p: Optional[float] = field(default=1.0)
typical_p: Optional[float] = field(default=1.0)
diversity_penalty: Optional[float] = field(default=0.0)
repetition_penalty: Optional[float] = field(default=1.0)
length_penalty: Optional[float] = field(default=1.0)
no_repeat_ngram_size: Optional[int] = field(default=0)
def find_all_linear_names(args, model):
cls = GeneralQuantLinear if not(args.full_finetune) else torch.nn.Linear
lora_module_names = set()
for name, module in model.named_modules():
if isinstance(module, cls):
names = name.split('.')
lora_module_names.add(names[0] if len(names) == 1 else names[-1])
if 'lm_head' in lora_module_names: # needed for 16-bit
lora_module_names.remove('lm_head')
return list(lora_module_names)
class SavePeftModelCallback(transformers.TrainerCallback):
def save_model(self, args, state, kwargs):
print('Saving PEFT checkpoint...')
if state.best_model_checkpoint is not None:
checkpoint_folder = os.path.join(state.best_model_checkpoint, "adapter_model")
else:
checkpoint_folder = os.path.join(args.output_dir, f"{PREFIX_CHECKPOINT_DIR}-{state.global_step}")
peft_model_path = os.path.join(checkpoint_folder, "adapter_model")
kwargs["model"].save_pretrained(peft_model_path)
pytorch_model_path = os.path.join(checkpoint_folder, "pytorch_model.bin")
if os.path.exists(pytorch_model_path):
os.remove(pytorch_model_path)
def on_save(self, args, state, control, **kwargs):
self.save_model(args, state, kwargs)
return control
def on_train_end(self, args, state, control, **kwargs):
def touch(fname, times=None):
with open(fname, 'a'):
os.utime(fname, times)
touch(join(args.output_dir, 'completed'))
self.save_model(args, state, kwargs)
def get_accelerate_model(args, checkpoint_dir):
n_gpus = torch.cuda.device_count()
max_memory = f'{args.max_memory_MB}MB'
max_memory = {i: max_memory for i in range(n_gpus)}
if args.full_finetune: assert args.bits in [16, 32]
print(f'loading base model {args.model_path}...')
model = AutoGPTQForCausalLM.from_quantized(
args.model_path,
device_map='auto',
max_memory=max_memory,
trust_remote_code=args.trust_remote_code,
inject_fused_attention = False,
inject_fused_mlp = False,
use_triton=True,
warmup_triton=False,
trainable=True
)
model.model.quantize_config = model.quantize_config
model.train()
setattr(model, 'model_parallel', True)
setattr(model, 'is_parallelizable', True)
#modules = find_all_linear_names(args, model)
model.config.torch_dtype=(torch.float32 if args.fp16 else (torch.bfloat16 if args.bf16 else torch.float32))
if not args.full_finetune:
model = prepare_model_for_int8_training(model, use_gradient_checkpointing=args.gradient_checkpointing)
if args.gradient_checkpointing:
model.gradient_checkpointing_enable()
config = GPTQLoraConfig(
r=args.lora_r,
lora_alpha=args.lora_alpha,
#target_modules=modules,
lora_dropout=args.lora_dropout,
bias="none",
task_type="CAUSAL_LM",
)
if not args.full_finetune:
if checkpoint_dir is not None:
print("Loading adapters from checkpoint.")
model = PeftModel.from_pretrained(model, join(checkpoint_dir, 'adapter_model'))
for name, p in model.named_parameters():
if 'lora' in name:
print(name, p.sum())
else:
print(f'adding LoRA modules...')
model = get_gptq_peft_model(model, config, auto_find_all_linears=True, train_mode=True)
if args.gradient_checkpointing:
if hasattr(model, "enable_input_require_grads"):
model.enable_input_require_grads()
else:
def make_inputs_require_grad(module, input, output):
output.requires_grad_(True)
model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)
for name, module in model.named_modules():
if isinstance(module, LoraLayer):
if args.bf16:
module = module.to(torch.bfloat16)
if 'norm' in name:
module = module.to(torch.float32)
if 'lm_head' in name or 'embed_tokens' in name:
if hasattr(module, 'weight'):
if args.bf16 and module.weight.dtype == torch.float32:
module = module.to(torch.bfloat16)
return model
def print_trainable_parameters(args, model):
"""
Prints the number of trainable parameters in the model.
"""
trainable_params = 0
all_param = 0
for _, param in model.named_parameters():
all_param += param.numel()
if param.requires_grad:
trainable_params += param.numel()
try:
trainable_params /= (32//model.quantize_config.bits)
except:
pass
print(f"trainable params: {trainable_params} || all params: {all_param} || trainable: {100 * trainable_params / all_param}")
def smart_tokenizer_and_embedding_resize(
special_tokens_dict: Dict,
tokenizer: transformers.PreTrainedTokenizer,
model: transformers.PreTrainedModel,
):
"""Resize tokenizer and embedding.
Note: This is the unoptimized version that may make your embedding size not be divisible by 64.
"""
num_new_tokens = tokenizer.add_special_tokens(special_tokens_dict)
model.resize_token_embeddings(len(tokenizer))
if num_new_tokens > 0:
input_embeddings = model.get_input_embeddings().weight.data
output_embeddings = model.get_output_embeddings().weight.data
input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True)
output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True)
input_embeddings[-num_new_tokens:] = input_embeddings_avg
output_embeddings[-num_new_tokens:] = output_embeddings_avg
@dataclass
class DataCollatorForCausalLM(object):
tokenizer: transformers.PreTrainedTokenizer
source_max_len: int
target_max_len: int
train_on_source: bool
predict_with_generate: bool
def __call__(self, instances: Sequence[Dict]) -> Dict[str, torch.Tensor]:
# Extract elements
sources = [example['input'] for example in instances]
targets = [f"{example['output']}{self.tokenizer.eos_token}" for example in instances]
# Tokenize
tokenized_sources_with_prompt = self.tokenizer(
sources,
max_length=self.source_max_len,
truncation=True,
)
tokenized_targets = self.tokenizer(
targets,
max_length=self.target_max_len,
truncation=True,
add_special_tokens=False,
)
# Build the input and labels for causal LM
input_ids = []
labels = []
for tokenized_source, tokenized_target in zip(
tokenized_sources_with_prompt['input_ids'],
tokenized_targets['input_ids']
):
if not self.predict_with_generate:
input_ids.append(torch.tensor(tokenized_source + tokenized_target))
if not self.train_on_source:
labels.append(
torch.tensor([IGNORE_INDEX for _ in range(len(tokenized_source))] + copy.deepcopy(tokenized_target))
)
else:
labels.append(torch.tensor(copy.deepcopy(tokenized_source + tokenized_target)))
else:
input_ids.append(torch.tensor(tokenized_source))
# Apply padding
input_ids = pad_sequence(input_ids, batch_first=True, padding_value=self.tokenizer.pad_token_id)
labels = pad_sequence(labels, batch_first=True, padding_value=IGNORE_INDEX) if not self.predict_with_generate else None
data_dict = {
'input_ids': input_ids,
'attention_mask':input_ids.ne(self.tokenizer.pad_token_id),
}
if labels is not None:
data_dict['labels'] = labels
return data_dict
def extract_unnatural_instructions_data(examples, extract_reformulations=False):
out = {
'input': [],
'output': [],
}
for example_instances in examples['instances']:
for instance in example_instances:
out['input'].append(instance['instruction_with_input'])
out['output'].append(instance['output'])
if extract_reformulations:
for example_reformulations in examples['reformulations']:
if example_reformulations is not None:
for instance in example_reformulations:
out['input'].append(instance['instruction_with_input'])
out['output'].append(instance['output'])
return out
PROMPT_DICT = {
"prompt_input": (
"Below is an instruction that describes a task, paired with an input that provides further context. "
"Write a response that appropriately completes the request.\n\n"
"### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response: "
),
"prompt_no_input": (
"Below is an instruction that describes a task. "
"Write a response that appropriately completes the request.\n\n"
"### Instruction:\n{instruction}\n\n### Response: "
),
}
def extract_alpaca_dataset(example):
if example.get("input", "") != "":
prompt_format = PROMPT_DICT["prompt_input"]
else:
prompt_format = PROMPT_DICT["prompt_no_input"]
return {'input': prompt_format.format(**example)}
def make_data_module(tokenizer: transformers.PreTrainedTokenizer, args) -> Dict:
"""
Make dataset and collator for supervised fine-tuning.
Datasets are expected to have the following columns: { `input`, `output` }
Available datasets to be selected with `dataset` argument:
- alpaca, 52002 examples
- alpaca cleaned, 51942 examples
- chip2 (OIG), 210289 examples
- self-instruct, 82612 examples
- hh-rlhf (Anthropic), 160800 examples
- longform, 23.7k examples
Coming soon:
- unnatural instructions core, 66010 examples
- unnatural instructions full, 240670 examples
- alpaca-gpt4, 52002 examples
- unnatural-instructions-gpt4, 9000 examples
- oa-rlhf (OpenAssistant) primary message tree only, 9209 examples
- oa-rlhf-assistant (OpenAssistant) all assistant replies with ranking
- supernatural-instructions, 69624 examples (same as paper with 100 ex/task more can be used)
- flan (FLAN v2), up to 20M examples available
Not Available:
- vicuna, not released at the moment.
"""
# Load dataset.
# Alpaca
if args.dataset == 'alpaca':
dataset = load_dataset("tatsu-lab/alpaca")
dataset = dataset.map(extract_alpaca_dataset, remove_columns=['instruction'])
# Alpaca clean
elif args.dataset == 'alpaca-clean':
dataset = load_dataset("yahma/alpaca-cleaned")
dataset = dataset.map(extract_alpaca_dataset, remove_columns=['instruction'])
# Chip2
elif args.dataset == 'chip2':
dataset = load_dataset("laion/OIG", data_files='unified_chip2.jsonl')
dataset = dataset.map(lambda x: {
'input': x['text'].split('\n<bot>: ')[0].replace('<human>: ', ''),
'output': x['text'].split('\n<bot>: ')[1],
}, remove_columns=['text', 'metadata'])
# Self Instruct
elif args.dataset == 'self-instruct':
dataset = load_dataset("yizhongw/self_instruct", name='self_instruct')
for old, new in [["prompt", "input"], ["completion", "output"]]:
dataset = dataset.rename_column(old, new)
# Anthropic rlhf
elif args.dataset == 'hh-rlhf':
dataset = load_dataset("Anthropic/hh-rlhf")
dataset = dataset.map(lambda x: {
'input': '',
'output': x['chosen']
}, remove_columns=['chosen', 'rejected'])
# LongForm
elif args.dataset == 'longform':
dataset = load_dataset("akoksal/LongForm")
elif args.dataset == 'vicuna':
raise NotImplementedError("Vicuna data was not released.")
else:
raise NotImplementedError(f"Dataset {args.dataset} not implemented yet.")
# Split train/eval, reduce size
if args.do_eval or args.do_predict:
if 'eval' in dataset:
eval_dataset = dataset['eval']
else:
print('Splitting train dataset in train and validation according to `eval_dataset_size`')
dataset = dataset["train"].train_test_split(
test_size=args.eval_dataset_size, shuffle=True, seed=42
)
eval_dataset = dataset['test']
if args.max_eval_samples is not None and len(eval_dataset) > args.max_eval_samples:
eval_dataset = eval_dataset.select(range(args.max_eval_samples))
if args.group_by_length:
eval_dataset = eval_dataset.map(lambda x: {'length': len(x['input']) + len(x['output'])})
if args.do_train:
train_dataset = dataset['train']
if args.max_train_samples is not None and len(train_dataset) > args.max_train_samples:
train_dataset = train_dataset.select(range(args.max_train_samples))
if args.group_by_length:
train_dataset = train_dataset.map(lambda x: {'length': len(x['input']) + len(x['output'])})
data_collator = DataCollatorForCausalLM(
tokenizer=tokenizer,
source_max_len=args.source_max_len,
target_max_len=args.target_max_len,
train_on_source=args.train_on_source,
predict_with_generate=args.predict_with_generate,
)
return dict(
train_dataset=train_dataset if args.do_train else None,
eval_dataset=eval_dataset if args.do_eval else None,
predict_dataset=eval_dataset if args.do_predict else None,
data_collator=data_collator
)
def get_last_checkpoint(checkpoint_dir):
if isdir(checkpoint_dir):
is_completed = exists(join(checkpoint_dir, 'completed'))
if is_completed: return None, True # already finished
max_step = 0
for filename in os.listdir(checkpoint_dir):
if isdir(join(checkpoint_dir, filename)) and filename.startswith('checkpoint'):
max_step = max(max_step, int(filename.replace('checkpoint-', '')))
if max_step == 0: return None, is_completed # training started, but no checkpoint
checkpoint_dir = join(checkpoint_dir, f'checkpoint-{max_step}')
print(f"Found a previous checkpoint at: {checkpoint_dir}")
return checkpoint_dir, is_completed # checkpoint found!
return None, False # first training
def train():
hfparser = transformers.HfArgumentParser((
ModelArguments, DataArguments, TrainingArguments, GenerationArguments
))
model_args, data_args, training_args, generation_args, extra_args = \
hfparser.parse_args_into_dataclasses(return_remaining_strings=True)
training_args.generation_config = transformers.GenerationConfig(**vars(generation_args))
args = argparse.Namespace(
**vars(model_args), **vars(data_args), **vars(training_args)
)
checkpoint_dir, completed_training = get_last_checkpoint(args.output_dir)
if completed_training:
print('Detected that training was already completed!')
model = get_accelerate_model(args, checkpoint_dir)
training_args.skip_loading_checkpoint_weights=True
resume_from_checkpoint = checkpoint_dir
if resume_from_checkpoint:
# Check the available weights and load them
checkpoint_name = os.path.join(
checkpoint_dir, "pytorch_model.bin"
) # Full checkpoint
if not os.path.exists(checkpoint_name):
checkpoint_path = os.path.join(
checkpoint_dir, "adapter_model"
)
checkpoint_name = os.path.join(
checkpoint_path, "adapter_model.bin"
) # only LoRA model - LoRA config above has to fit
resume_from_checkpoint = (
False # So the trainer won't try loading its state
)
# The two files above have a different name depending on how they were saved, but are actually the same.
if os.path.exists(checkpoint_name):
print(f"Restarting from {checkpoint_name}")
adapters_weights = torch.load(checkpoint_name)
set_peft_model_state_dict(model, adapters_weights)
else:
print(f"Checkpoint {checkpoint_name} not found")
model.config.use_cache = False
print_trainable_parameters(args, model)
print('loaded model')
set_seed(args.seed)
# Tokenizer
tokenizer = AutoTokenizer.from_pretrained(
args.model_path,
cache_dir=args.cache_dir,
padding_side="right",
use_fast=True,
)
if tokenizer.pad_token is None:
smart_tokenizer_and_embedding_resize(
special_tokens_dict=dict(pad_token=DEFAULT_PAD_TOKEN),
tokenizer=tokenizer,
model=model,
)
if isinstance(tokenizer, LlamaTokenizerFast):
# LLaMA tokenizer may not have correct special tokens set.
# Check and add them if missing to prevent them from being parsed into different tokens.
# Note that these are present in the vocabulary.
# Note also that `model.config.pad_token_id` is 0 which corresponds to `<unk>` token.
tokenizer.add_special_tokens(
{
"eos_token": tokenizer.convert_ids_to_tokens(model.config.eos_token_id),
"bos_token": tokenizer.convert_ids_to_tokens(model.config.bos_token_id),
"unk_token": tokenizer.convert_ids_to_tokens(model.config.pad_token_id),
}
)
data_module = make_data_module(tokenizer=tokenizer, args=args)
trainer = Seq2SeqTrainer(
model=model,
tokenizer=tokenizer,
args=training_args,
**{k:v for k,v in data_module.items() if k != 'predict_dataset'},
)
# Callbacks
if not args.full_finetune:
trainer.add_callback(SavePeftModelCallback)
if args.do_mmlu_eval:
if args.mmlu_dataset == 'mmlu-zs':
mmlu_dataset = load_dataset("json", data_files={
'eval': 'data/mmlu/zero_shot_mmlu_val.json',
'test': 'data/mmlu/zero_shot_mmlu_test.json',
})
mmlu_dataset = mmlu_dataset.remove_columns('subject')
# MMLU Five-shot (Eval/Test only)
elif args.mmlu_dataset == 'mmlu' or args.mmlu_dataset == 'mmlu-fs':
mmlu_dataset = load_dataset("json", data_files={
'eval': 'data/mmlu/five_shot_mmlu_val.json',
'test': 'data/mmlu/five_shot_mmlu_test.json',
})
# mmlu_dataset = mmlu_dataset.remove_columns('subject')
mmlu_dataset = mmlu_dataset[args.mmlu_split]
if args.max_mmlu_samples is not None:
mmlu_dataset = mmlu_dataset.select(range(args.max_mmlu_samples))
abcd_idx = [
tokenizer("A", add_special_tokens=False).input_ids[0],
tokenizer("B", add_special_tokens=False).input_ids[0],
tokenizer("C", add_special_tokens=False).input_ids[0],
tokenizer("D", add_special_tokens=False).input_ids[0],
]
accuracy = evaluate.load("accuracy")
class MMLUEvalCallback(transformers.TrainerCallback):
def on_evaluate(self, args, state, control, model, **kwargs):
data_loader = trainer.get_eval_dataloader(mmlu_dataset)
source_max_len = trainer.data_collator.source_max_len
trainer.data_collator.source_max_len = args.mmlu_source_max_len
trainer.model.eval()
preds, refs = [], []
loss_mmlu = 0
for batch in tqdm(data_loader, total=len(data_loader)):
(loss, logits, labels) = trainer.prediction_step(trainer.model,batch,prediction_loss_only=False,)
# There are two tokens, the output, and eos token.
for i, logit in enumerate(logits):
label_non_zero_id = (batch['labels'][i] != -100).nonzero()[0][0]
logit_abcd = logit[label_non_zero_id-1][abcd_idx]
preds.append(torch.argmax(logit_abcd).item())
labels = labels[labels != IGNORE_INDEX].view(-1, 2)[:,0]
for label in labels.tolist():
if label in abcd_idx:
refs += [abcd_idx.index(label)]
loss_mmlu += loss.item()
# Extract results by subject.
results = {'mmlu_loss':loss_mmlu/len(data_loader)}
subject = mmlu_dataset['subject']
subjects = {s:{'refs':[], 'preds':[]} for s in set(subject)}
for s,p,r in zip(subject, preds, refs):
subjects[s]['preds'].append(p)
subjects[s]['refs'].append(r)
subject_scores = []
for subject in subjects:
subject_score = accuracy.compute(
references=subjects[subject]['refs'],
predictions=subjects[subject]['preds']
)['accuracy']
results[f'mmlu_{args.mmlu_split}_accuracy_{subject}'] = subject_score
subject_scores.append(subject_score)
results[f'mmlu_{args.mmlu_split}_accuracy'] = np.mean(subject_scores)
trainer.log(results)
trainer.data_collator.source_max_len = source_max_len
trainer.add_callback(MMLUEvalCallback)
# Verifying the datatypes.
dtypes = {}
for _, p in model.named_parameters():
dtype = p.dtype
if dtype not in dtypes: dtypes[dtype] = 0
dtypes[dtype] += p.numel()
total = 0
for k, v in dtypes.items(): total+= v
for k, v in dtypes.items():
print(k, v, v/total)
all_metrics = {"run_name": args.run_name}
# Training
if args.do_train:
train_result = trainer.train(resume_from_checkpoint=resume_from_checkpoint)
metrics = train_result.metrics
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
all_metrics.update(metrics)
# Evaluation
if args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate(metric_key_prefix="eval")
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
all_metrics.update(metrics)
# Prediction
if args.do_predict:
logger.info("*** Predict ***")
prediction_output = trainer.predict(test_dataset=data_module['predict_dataset'],metric_key_prefix="predict")
prediction_metrics = prediction_output.metrics
predictions = prediction_output.predictions
predictions = np.where(predictions != -100, predictions, tokenizer.pad_token_id)
predictions = tokenizer.batch_decode(
predictions, skip_special_tokens=True, clean_up_tokenization_spaces=True
)
with open(os.path.join(args.output_dir, 'predictions.jsonl'), 'w') as fout:
for i, example in enumerate(data_module['predict_dataset']):
example['prediction_with_input'] = predictions[i].strip()
example['prediction'] = predictions[i].replace(example['input'], '').strip()
fout.write(json.dumps(example) + '\n')
print(prediction_metrics)
trainer.log_metrics("predict", prediction_metrics)
trainer.save_metrics("predict", prediction_metrics)
all_metrics.update(prediction_metrics)
if (args.do_train or args.do_eval or args.do_predict):
with open(os.path.join(args.output_dir, "metrics.json"), "w") as fout:
fout.write(json.dumps(all_metrics))
if __name__ == "__main__":
train()
================================================
FILE: requirements.txt
================================================
bert-score==0.3.13
evaluate==0.4.0
rouge-score==0.1.2
scikit-learn==1.2.2
sentencepiece==0.1.99
wandb==0.15.2
transformers==4.31.0
peft==0.4.0
accelerate==0.21.0
gitextract_jckgmwf6/ ├── LICENSE ├── README.md ├── environment.yaml ├── merge.py ├── peft_utils.py ├── qalora.py └── requirements.txt
SYMBOL INDEX (48 symbols across 2 files)
FILE: peft_utils.py
class GPTQLoraConfig (line 22) | class GPTQLoraConfig(LoraConfig):
class GPTQLoraLinear (line 27) | class GPTQLoraLinear(torch.nn.Linear, LoraLayer):
method __init__ (line 28) | def __init__(
method reset_lora_parameters (line 56) | def reset_lora_parameters(self, adapter_name):
method merge (line 61) | def merge(self):
method unmerge (line 64) | def unmerge(self):
method forward (line 67) | def forward(self, x: torch.Tensor):
class GPTQLoraModel (line 94) | class GPTQLoraModel(LoraModel):
method _find_and_replace (line 95) | def _find_and_replace(self, adapter_name):
method _replace_module (line 155) | def _replace_module(self, parent_module, child_name, new_module, old_m...
method merge_adapter (line 172) | def merge_adapter(self):
method unmerge_adapter (line 175) | def unmerge_adapter(self):
method merge_and_unload (line 178) | def merge_and_unload(self):
class GPTQAdaLoraConfig (line 182) | class GPTQAdaLoraConfig(AdaLoraConfig):
class GPTQSVDLinear (line 187) | class GPTQSVDLinear(torch.nn.Linear, AdaLoraLayer):
method __init__ (line 188) | def __init__(
method merge (line 215) | def merge(self):
method unmerge (line 218) | def unmerge(self):
method forward (line 221) | def forward(self, x: torch.Tensor):
class GPTQAdaLoraModel (line 244) | class GPTQAdaLoraModel(AdaLoraModel):
method _find_and_replace (line 245) | def _find_and_replace(self, adapter_name):
method _replace_module (line 297) | def _replace_module(self, parent_module, child_name, new_module, old_m...
method merge_adapter (line 305) | def merge_adapter(self):
method unmerge_adapter (line 308) | def unmerge_adapter(self):
method merge_and_unload (line 311) | def merge_and_unload(self):
function find_all_linear_names (line 315) | def find_all_linear_names(model: BaseGPTQForCausalLM, ignore: Optional[L...
function hijack_peft_mappings (line 331) | def hijack_peft_mappings():
function get_gptq_peft_model (line 352) | def get_gptq_peft_model(
FILE: qalora.py
function prepare_model_for_int8_training (line 52) | def prepare_model_for_int8_training(model, use_gradient_checkpointing=Tr...
class ModelArguments (line 88) | class ModelArguments:
class DataArguments (line 98) | class DataArguments:
class TrainingArguments (line 130) | class TrainingArguments(transformers.Seq2SeqTrainingArguments):
class GenerationArguments (line 206) | class GenerationArguments:
function find_all_linear_names (line 237) | def find_all_linear_names(args, model):
class SavePeftModelCallback (line 251) | class SavePeftModelCallback(transformers.TrainerCallback):
method save_model (line 252) | def save_model(self, args, state, kwargs):
method on_save (line 266) | def on_save(self, args, state, control, **kwargs):
method on_train_end (line 270) | def on_train_end(self, args, state, control, **kwargs):
function get_accelerate_model (line 278) | def get_accelerate_model(args, checkpoint_dir):
function print_trainable_parameters (line 352) | def print_trainable_parameters(args, model):
function smart_tokenizer_and_embedding_resize (line 368) | def smart_tokenizer_and_embedding_resize(
class DataCollatorForCausalLM (line 391) | class DataCollatorForCausalLM(object):
method __call__ (line 398) | def __call__(self, instances: Sequence[Dict]) -> Dict[str, torch.Tensor]:
function extract_unnatural_instructions_data (line 442) | def extract_unnatural_instructions_data(examples, extract_reformulations...
function extract_alpaca_dataset (line 472) | def extract_alpaca_dataset(example):
function make_data_module (line 479) | def make_data_module(tokenizer: transformers.PreTrainedTokenizer, args) ...
function get_last_checkpoint (line 576) | def get_last_checkpoint(checkpoint_dir):
function train (line 590) | def train():
Condensed preview — 7 files, each showing path, character count, and a content snippet. Download the .json file or copy for the full structured content (62K chars).
[
{
"path": "LICENSE",
"chars": 1064,
"preview": "MIT License\n\nCopyright (c) 2023 YuhuiXu\n\nPermission is hereby granted, free of charge, to any person obtaining a copy\nof"
},
{
"path": "README.md",
"chars": 3421,
"preview": "# QA-LoRA\n\nQA-LoRA has been accepted by ICLR 2024!\n\nThis repository provides the official PyTorch implementation of [QA-"
},
{
"path": "environment.yaml",
"chars": 2567,
"preview": "name: alpaca\nchannels:\n - defaults\ndependencies:\n - _libgcc_mutex=0.1=main\n - _openmp_mutex=5.1=1_gnu\n - ca-certific"
},
{
"path": "merge.py",
"chars": 611,
"preview": "import torch\r\nmodel_path = 'path of the quantized model'\r\nlora_path = 'path of the saved LoRA adapters'\r\nmerged_path = '"
},
{
"path": "peft_utils.py",
"chars": 18465,
"preview": "import warnings\nimport re\nfrom contextlib import contextmanager\nfrom dataclasses import asdict\nfrom enum import Enum\nfro"
},
{
"path": "qalora.py",
"chars": 33407,
"preview": "# This source code is licensed under the MIT license found in the\n# LICENSE file in the root directory of this source tr"
},
{
"path": "requirements.txt",
"chars": 161,
"preview": "bert-score==0.3.13\nevaluate==0.4.0\nrouge-score==0.1.2\nscikit-learn==1.2.2\nsentencepiece==0.1.99\nwandb==0.15.2\ntransforme"
}
]
About this extraction
This page contains the full source code of the yuhuixu1993/qa-lora GitHub repository, extracted and formatted as plain text for AI agents and large language models (LLMs). The extraction includes 7 files (58.3 KB), approximately 14.3k tokens, and a symbol index with 48 extracted functions, classes, methods, constants, and types. Use this with OpenClaw, Claude, ChatGPT, Cursor, Windsurf, or any other AI tool that accepts text input. You can copy the full output to your clipboard or download it as a .txt file.
Extracted by GitExtract — free GitHub repo to text converter for AI. Built by Nikandr Surkov.