[DO NOT MERGE] 实现与PaddleFormers对齐的Qwen2

[JunnYu]

Before submitting

• Lint code. If there are lint issues, please format the code first.

# Install and register `pre-commit` in the project folder
pip install pre-commit && pre-commit install

# Process previous code files separately
pre-commit run --file XXXX.py

• Add test cases into tests folder. If there are codecov issues, please add tests cases first.

PR types

PR changes

Description

为了对齐PaddleFormers的Qwen2.5，特此进行了修改。

• 对齐rope旋转位置编码精度，提升精度采用FP32进行计算。
• 对齐fuse配置逻辑，开启fuse_swiglu, 开启 fuse_rms_norm
• 对齐fuse rms norm的实现，均采用框架实现（关闭fast math编译）
• 对齐LM Head的权重shape
  • paddleformers的实现：logits = paddle.matmul(x, weight, transpose_y=True) weight的shape[vocab_size, hidden_size]
  • paddlenlp的实现：
    • logits = paddle.matmul(x, weight, transpose_y=False) weight的shape[hidden_size, vocab_size]
    • [实现1] 如果将 logits = paddle.matmul(x, weight.t(), transpose_y=True) weight的shape[hidden_size, vocab_size] 这样实现，前向精度能保持一致，但是反向的时候在训练过程中如果开启main_grad后，梯度不一致。
    • [实现2] 因此，必须修改PaddleNLP的组网，使得paddlenlp的weight的shape与paddleformers要一模一样！

[paddle-bot]

Thanks for your contribution!