add common models and common params unit test for llm predictor.

[aooxin]

PR types

Others

PR changes

Others

Description

添加llm推理中常见模型和常见参数的单测
在A30单卡暂时耗时如下：
15 passed in 1507.02s (0:18:27)

覆盖下面的模型

	模型名称
1	Qwen2.5-7B-Instruct-tiny-nhl2
2	Qwen1.5-MoE-A2.7B-Chat-tiny-nhl2
3	Llama-2-7b-chat-tiny-nhl2
4	Meta-Llama-3.1-8B-Instruct-tiny-nhl2

Qwen2.5-1.5B-Instruct在参数测试中覆盖
70B和72B的在多卡中测试

覆盖下面的参数

两个 model：Qwen2.5-1.5B-Instruct和meta-llama/Llama-3.2-3B-Instruct

参数名称	值
use_fake_parameter	True
inference_model	False
append_attn	True

覆盖分布式多卡推理场景

	模型名称
2	Qwen2.5-72B-Instruct-tiny-nhl2
3	Llama-2-70b-chat-tiny-nhl2
4	Meta-Llama-3.1-70B-Instruct-tiny-nhl2

[paddle-bot]

Thanks for your contribution!

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 52.77%. Comparing base (3aa9f4c) to head (863ff0b).

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #9711      +/-   ##
===========================================
- Coverage    53.30%   52.77%   -0.53%     
===========================================
  Files          724      718       -6     
  Lines       118271   112449    -5822     
===========================================
- Hits         63039    59350    -3689     
+ Misses       55232    53099    -2133     

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