[LLM]Support Gemma model

[Southpika]

PR types

New features

PR changes

Models

Description

添加gemma 2b/gemma 7b模型

[paddle-bot]

Thanks for your contribution!

[codecov]

Codecov Report

Attention: Patch coverage is 66.56315% with 323 lines in your changes are missing coverage. Please review.

Project coverage is 55.26%. Comparing base (2273ee7) to head (2f8b5d8).
Report is 2 commits behind head on develop.

❗ Current head 2f8b5d8 differs from pull request most recent head bdb18ef. Consider uploading reports for the commit bdb18ef to get more accurate results

Files	Patch %	Lines
paddlenlp/transformers/gemma/modeling.py	68.62%	219 Missing ⚠️
paddlenlp/transformers/gemma/fast_tokenizer.py	0.00%	70 Missing ⚠️
paddlenlp/transformers/gemma/tokenizer.py	76.42%	33 Missing ⚠️
paddlenlp/transformers/model_utils.py	87.50%	1 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #8082      +/-   ##
===========================================
+ Coverage    55.15%   55.26%   +0.11%     
===========================================
  Files          601      606       +5     
  Lines        91764    92728     +964     
===========================================
+ Hits         50611    51250     +639     
- Misses       41153    41478     +325     

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

[Southpika]

zero padding曲线基本对齐

[Southpika]

Gemma-7b最终sft曲线比较，蓝色为tp，绿色为pp，紫色为sharding

[wawltor]

LGTM