Release/3.0 beta3 cherry-pick (#9554)

* [Fix] Update requirements.txt (#9514)

* Update README.md

Author: ZHUI

README.md

@@ -156,7 +156,7 @@
 ### pip 安装
 
 ```shell
-pip install --upgrade paddlenlp==3.0.0b2
+pip install --upgrade paddlenlp==3.0.0b3
 ```
 
 或者可通过以下命令安装最新 develop 分支代码：

llm/docs/finetune.md

@@ -36,7 +36,7 @@
 ### 3.1 环境准备
 
 - PaddlePaddle 3.0-beta
-- PaddleNLP   3.0.0b2
+- PaddleNLP   3.0.0b3
 - PaddleSlim develop
 
 git clone 代码到本地，即可开始。

llm/run_finetune.py

@@ -79,7 +79,18 @@
 flash_mask_support_list = [LlamaForCausalLM, LlamaForCausalLMPipe, Qwen2ForCausalLM, Qwen2ForCausalLMPipe]
 
 
+def paddlenlp_verison_check():
+    import paddlenlp
+    from paddlenlp.utils.tools import compare_version
+
+    if not compare_version(paddlenlp.__version__, "3.0.0.b2"):
+        raise ValueError(
+            "This scripts require paddlenlp >= 3.0.0b3, please reinstall: pip install paddlenlp >= 3.0.0b3 "
+        )
+
+
 def main():
+    paddlenlp_verison_check()
     parser = PdArgumentParser((GenerateArgument, ModelConfig, ReftArgument, DataConfig, SFTConfig))
     if len(sys.argv) >= 2 and sys.argv[1].endswith(".json"):
         gen_args, model_args, reft_args, data_args, training_args = parser.parse_json_file_and_cmd_lines()

paddlenlp/__init__.py

@@ -20,7 +20,7 @@
 
 # this version is used for develop and test.
 # release version will be added fixed version by setup.py.
-__version__ = "3.0.0b2.post"
+__version__ = "3.0.0b3.post"
 if os.getenv(PADDLENLP_STABLE_VERSION):
     __version__ = __version__.replace(".post", "")
 else:

requirements.txt

@@ -27,4 +27,4 @@ jinja2
 regex
 numpy<=1.26.4
 tiktoken
-tokenizers
+tokenizers<0.21

setup.py

@@ -112,7 +112,7 @@ def show():
 
 
 # only use this file to contral the version
-__version__ = "3.0.0b2.post"
+__version__ = "3.0.0b3.post"
 if os.getenv(PADDLENLP_STABLE_VERSION):
     __version__ = __version__.replace(".post", "")
 else: