Adding LoKrModel Class to paddle.peft library #9269

WhuanY · 2024-10-15T06:23:57Z

PR types

New features

PR changes

Others

Description

Adding LoKrModel, LoKrLinear and LoKrConfig to support a new lora-like adapter. Current implementation only supports contains Linear Modules. Motivation and discussion on such PR issue is at: #9226

Please provide suggestions on the current implementation!

paddle-bot · 2024-10-15T06:24:02Z

Thanks for your contribution!

codecov · 2024-10-15T06:55:02Z

Codecov Report

Attention: Patch coverage is 80.61798% with 69 lines in your changes missing coverage. Please review.

Project coverage is 53.03%. Comparing base (f5ca96e) to head (ec91282).
Report is 4 commits behind head on develop.

Files with missing lines	Patch %	Lines
paddlenlp/peft/lokr/lokr_model.py	80.10%	38 Missing ⚠️
paddlenlp/peft/lokr/lokr_layers.py	71.28%	29 Missing ⚠️
paddlenlp/trainer/trainer.py	33.33%	2 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #9269      +/-   ##
===========================================
- Coverage    53.10%   53.03%   -0.08%     
===========================================
  Files          692      694       +2     
  Lines       110570   110254     -316     
===========================================
- Hits         58715    58470     -245     
+ Misses       51855    51784      -71

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

WhuanY · 2024-10-16T02:44:43Z

@DesmonDay

按照建议我已经提交了只有Linear Layer的LoKr实现供你们整体查看。麻烦有时间审阅下并给出需要修改的意见。

greycooker · 2024-10-16T06:22:38Z

好的收到，我们这边会尽快review

lugimzzz · 2024-10-16T06:21:03Z

paddlenlp/peft/lokr/lokr_envs.py

+
+# This module is set to be in alignment with code design paradiam of ...utils.env
+
+LOKR_WEIGHTS_NAME = "lokr_model_state.pdparams"


统一写到这里https://github.com/PaddlePaddle/PaddleNLP/blob/develop/paddlenlp/utils/env.py

lugimzzz · 2024-10-16T09:02:25Z

paddlenlp/peft/lokr/__init__.py

@@ -0,0 +1,19 @@
+# Copyright 2023-present the HuggingFace Inc. team.


copyright注意修改

paddlenlp/peft/lokr/__init__.py

lugimzzz · 2024-10-17T04:18:16Z

paddlenlp/peft/lokr/lokr_model.py

+    def add_lora_split_mapping(self, module_name, is_column=False):
+        self.lora_split_mapping[module_name] = is_column
+
+    def _get_tensor_parallel_mappings(self, config, is_split=True):


没有使用tensor_parallel 和pipeline parallel先把没用到的相关的逻辑删掉

lugimzzz · 2024-10-17T04:19:30Z

paddlenlp/peft/lokr/lokr_model.py

+            self.quantized = True
+        if lokr_module is None:
+            raise ValueError("LoKr strategy only supports paddle.nn.Linear right now")
+        if getattr(lokr_module, "quant_weight", None) is not None:


quant相关逻辑没用到的也先删掉

paddlenlp/peft/lokr/lokr_model.py

WhuanY · 2024-10-17T05:06:29Z

辛苦了！第一次参与开源。可能错误较多，我会尽早修改问题，重新提交，供你们审阅。 Have a good day!

lugimzzz · 2024-10-22T06:51:37Z

辛苦了！第一次参与开源。可能错误较多，我会尽早修改问题，重新提交，供你们审阅。 Have a good day!

感谢对PaddleNLP的贡献，我们非常欢迎社区开发者参与到PaddleNLP的开发中来。我会在重新提交代码后尽快进行review，期待提交的代码能早日合入到项目中！

lugimzzz · 2024-10-29T08:22:22Z

可以再次review请告知我，我会尽快开始review

WhuanY · 2024-10-30T13:23:00Z

可以再次review请告知我，我会尽快开始review

按照要求我已经

去掉了暂时不涉及的并行逻辑；
增加了disable_lokr参数和相应办法
增加了test/peft/lokr_model.py，并通过了基本的测试；
根据test过程中发现的bug更新了部分LoKrLinear，包括重置初始化方式、更正前向传播Bug。

目前我可以想到的接下来可以做的是：

在unified_checkpoint中支持LoKrModel
增加该适配器的合并参数脚本
如有问题和改进方向请说明，辛苦了！

lugimzzz · 2024-11-01T06:37:03Z

llm/config/llama/lokr_argument.json

+    "model_name_or_path": "meta-llama/Meta-Llama-3-8B",
+    "dataset_name_or_path": "./data",
+    "output_dir": "./checkpoints/lokr_ckpts",
+    "lokr": true,


1.代码还未在https://github.com/PaddlePaddle/PaddleNLP/blob/develop/llm/run_finetune.py 文件中添加相应的代码，没有看到lokr设为true时执行的逻辑。
2.添加后请相应同步文档，lokr运行方式以及对应新增参数的解释https://github.com/PaddlePaddle/PaddleNLP/blob/develop/llm/docs/finetune.md
3.模仿lora添加llm单测 https://github.com/PaddlePaddle/PaddleNLP/blob/develop/tests/llm/test_lora.py
4.请参考vera和lora的脚本新增一个 merge_lokr_params.py https://github.com/PaddlePaddle/PaddleNLP/blob/develop/llm/tools/merge_vera_params.py

再次感谢对开源代码的贡献，代码库中LoKr算法实现没有问题，补充完大模型应用样例即可合入PaddleNLP

好的，收到！争取这周末完成上述四点，到时候我在远程仓库提交

WhuanY · 2024-11-16T13:48:11Z

可以再次review请告知我，我会尽快开始review

你好，应该可以开始review了，最近项目私下做了测试，通过功能跨平台精度对齐的任务证明已经没有算法错误

lugimzzz · 2024-11-20T11:00:19Z

关注一下单测覆盖率，PaddleNLP-CI报错看起来是网络问题，我rerun了。这两个问题解决就可以合入了

lugimzzz · 2024-11-22T03:20:17Z

需要解决一下冲突和单测覆盖率，即可合入 @WhuanY

WhuanY · 2024-11-27T12:50:11Z

你好，冲突和单测问题已经解决～看看还有什么需要修正的吗？ @lugimzzz

lugimzzz

LGTM！感谢对飞桨开源框架贡献❤️！

WhuanY · 2024-11-27T12:58:28Z

My pleasure❤️😄

passing pre-commit

21f222b

lugimzzz reviewed Oct 17, 2024

View reviewed changes

lugimzzz closed this Oct 22, 2024

lugimzzz reopened this Oct 22, 2024

WhuanY added 2 commits October 22, 2024 18:44

removing tp and pp logic for single gpu training

8054d8e

add disable_lokr attribute in lokr_layer

0e5a844

WhuanY mentioned this pull request Oct 22, 2024

[Question]: [关于开源贡献者]如何执行特定的单元测试？ #9302

Closed

refine comments

93f62e7

add lokr tests and modified layer bug

67aba6c

WhuanY added 3 commits October 30, 2024 21:26

add lokrtests

3e2703d

add lokrtests

aeaa619

add lokr_argument.json

7860fed

lugimzzz reviewed Nov 1, 2024

View reviewed changes

WhuanY mentioned this pull request Nov 6, 2024

[Question]: 集成测试代码出现ModuleNotFoundError. 安装失败，如何解决？ #9375

Closed

WhuanY added 3 commits November 6, 2024 15:29

add integration test, fix bugs based on tests.

98000cd

refactor lora_dim to lokr_dim

215beaa

no inference

4723293

WhuanY added 2 commits November 25, 2024 23:40

add more tests

0d72948

resolve merge conflict

f74a545

WhuanY and others added 3 commits November 26, 2024 21:37

Merge branch 'develop' into LoKrModel

d339509

add more randtests

056341f

pass isort check(maybe)

ec91282

lugimzzz approved these changes Nov 27, 2024

View reviewed changes

lugimzzz merged commit 3ef14dc into PaddlePaddle:develop Nov 27, 2024
10 of 12 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Adding LoKrModel Class to paddle.peft library #9269

Adding LoKrModel Class to paddle.peft library #9269

WhuanY commented Oct 15, 2024

paddle-bot bot commented Oct 15, 2024

codecov bot commented Oct 15, 2024 •

edited

Loading

WhuanY commented Oct 16, 2024

greycooker commented Oct 16, 2024

lugimzzz Oct 16, 2024

WhuanY Oct 22, 2024

lugimzzz Oct 16, 2024

WhuanY Oct 22, 2024

lugimzzz Oct 17, 2024

WhuanY Oct 22, 2024

lugimzzz Oct 17, 2024

WhuanY Oct 22, 2024

WhuanY commented Oct 17, 2024

lugimzzz commented Oct 22, 2024 •

edited

Loading

lugimzzz commented Oct 29, 2024

WhuanY commented Oct 30, 2024

lugimzzz Nov 1, 2024 •

edited

Loading

lugimzzz Nov 1, 2024

WhuanY Nov 1, 2024

WhuanY Nov 11, 2024

WhuanY Nov 20, 2024

WhuanY commented Nov 16, 2024

lugimzzz commented Nov 20, 2024

lugimzzz commented Nov 22, 2024 •

edited

Loading

WhuanY commented Nov 27, 2024

lugimzzz left a comment

WhuanY commented Nov 27, 2024


		# This module is set to be in alignment with code design paradiam of ...utils.env

		LOKR_WEIGHTS_NAME = "lokr_model_state.pdparams"

		@@ -0,0 +1,19 @@
		# Copyright 2023-present the HuggingFace Inc. team.

Adding LoKrModel Class to paddle.peft library #9269

Adding LoKrModel Class to paddle.peft library #9269

Conversation

WhuanY commented Oct 15, 2024

PR types

PR changes

Description

paddle-bot bot commented Oct 15, 2024

codecov bot commented Oct 15, 2024 • edited Loading

Codecov Report

WhuanY commented Oct 16, 2024

greycooker commented Oct 16, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

WhuanY commented Oct 17, 2024

lugimzzz commented Oct 22, 2024 • edited Loading

lugimzzz commented Oct 29, 2024

WhuanY commented Oct 30, 2024

lugimzzz Nov 1, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

WhuanY commented Nov 16, 2024

lugimzzz commented Nov 20, 2024

lugimzzz commented Nov 22, 2024 • edited Loading

WhuanY commented Nov 27, 2024

lugimzzz left a comment

Choose a reason for hiding this comment

WhuanY commented Nov 27, 2024

codecov bot commented Oct 15, 2024 •

edited

Loading

lugimzzz commented Oct 22, 2024 •

edited

Loading

lugimzzz Nov 1, 2024 •

edited

Loading

lugimzzz commented Nov 22, 2024 •

edited

Loading