[Infer] change block attention as default attention and remove normal attention in inference mode

[zeroRains]

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

Others

PR changes

Others

Description

在inference_model=1时，将block attention作为默认的attention并移除原本attention的使用，并在代码中移除不再使用的部分

1. 不设置--block_attn时，默认使用block attn
2. --append_attn=1时，使用append attn
3. --append_attn=0 --block_attn=0时，抛出异常

进度List :

• bloom

  测试指令：

  cd llm
  # block attn
  python ./predict/predictor.py --model_name_or_path bigscience/bloom-7b1 --dtype float16 --mode dynamic --decode_strategy greedy_search --inference_model 1 --block_attn 1 --append_attn 0
  
  # append attn
  python ./predict/predictor.py --model_name_or_path bigscience/bloom-7b1 --dtype float16 --mode dynamic --decode_strategy greedy_search --inference_model 1 --block_attn 0 --append_attn 1
  
  # throw exception
  python ./predict/predictor.py --model_name_or_path bigscience/bloom-7b1 --dtype float16 --mode dynamic --decode_strategy greedy_search --inference_model 1 --block_attn 0 --append_attn 0

  存在问题：
  append attn执行报错（develop也会）
  
  其他问题：bigscience/bloom-1b1模型的head_dim=96，block attn在head_dim=96时仍然有兼容性问题（develop也有）(已解决)
  
  [Infer] Add head_dim=96 dispatch for block attention Paddle#70763

• chatglm （没有block_attn的逻辑，保留原样）

• chatglm_v2
  测试指令：

  cd llm
  # block attn
  python ./predict/predictor.py --model_name_or_path THUDM/chatglm2-6b --dtype float16 --mode dynamic --decode_strategy greedy_search --inference_model 1 --block_attn 1 --append_attn 0
  
  # append attn
  python ./predict/predictor.py --model_name_or_path THUDM/chatglm2-6b --dtype float16 --mode dynamic --decode_strategy greedy_search --inference_model 1 --block_attn 0 --append_attn 1
  
  # throw exception
  python ./predict/predictor.py --model_name_or_path THUDM/chatglm2-6b --dtype float16 --mode dynamic --decode_strategy greedy_search --inference_model 1 --block_attn 0 --append_attn 0

  存在问题：
  append attn的计算会越界（develop也会）
  

• gpt（没有block_attn的逻辑，保留原样）

• llama

  测试指令：

  cd llm
  # block attn
  python ./predict/predictor.py --model_name_or_path meta-llama/Meta-Llama-3-8B-Instruct --dtype float16 --mode dynamic --decode_strategy greedy_search --inference_model 1 --block_attn 1 --append_attn 0
  
  # append attn
  python ./predict/predictor.py --model_name_or_path meta-llama/Meta-Llama-3-8B-Instruct --dtype float16 --mode dynamic --decode_strategy greedy_search --inference_model 1 --block_attn 0 --append_attn 1
  
  # block attn img2txt
  python ./predict/predictor.py --model_name_or_path meta-llama/Meta-Llama-3-8B-Instruct --dtype float16 --mode dynamic --decode_strategy greedy_search --inference_model 1 --block_attn 1 --append_attn 0 --model_type img2txt
  
  # append attn img2txt
  python ./predict/predictor.py --model_name_or_path meta-llama/Meta-Llama-3-8B-Instruct --dtype float16 --mode dynamic --decode_strategy greedy_search --inference_model 1 --block_attn 0 --append_attn 1 --model_type img2txt
  
  # throw exception
  python ./predict/predictor.py --model_name_or_path meta-llama/Meta-Llama-3-8B-Instruct --dtype float16 --mode dynamic --decode_strategy greedy_search --inference_model 1 --block_attn 0 --append_attn 0

• mixtral （本地环境爆显存了，需要帮忙测一下）
  测试指令：

  cd llm
  # block attn
  python ./predict/predictor.py --model_name_or_path mistralai/Mixtral-8x7B-Instruct-v0.1 --dtype float16 --mode dynamic --decode_strategy greedy_search --inference_model 1 --block_attn 1 --append_attn 0
  
  # append attn
  python ./predict/predictor.py --model_name_or_path mistralai/Mixtral-8x7B-Instruct-v0.1 --dtype float16 --mode dynamic --decode_strategy greedy_search --inference_model 1 --block_attn 0 --append_attn 1
  
  # throw exception
  python ./predict/predictor.py --model_name_or_path mistralai/Mixtral-8x7B-Instruct-v0.1 --dtype float16 --mode dynamic --decode_strategy greedy_search --inference_model 1 --block_attn 0 --append_attn 0

• opt（没有block_attn的逻辑，保留原样）

• qwen（没有block_attn的逻辑，保留原样）

• qwen2
  测试指令：

  cd llm
  # block attn
  python ./predict/predictor.py --model_name_or_path Qwen/Qwen2-1.5B-Instruct --dtype float16 --mode dynamic --decode_strategy greedy_search --inference_model 1 --block_attn 1 --append_attn 0
  
  # append attn
  python ./predict/predictor.py --model_name_or_path Qwen/Qwen2-1.5B-Instruct --dtype float16 --mode dynamic --decode_strategy greedy_search --inference_model 1 --block_attn 0 --append_attn 1
  
  # throw exception
  python ./predict/predictor.py --model_name_or_path Qwen/Qwen2-1.5B-Instruct --dtype float16 --mode dynamic --decode_strategy greedy_search --inference_model 1 --block_attn 0 --append_attn 0

• qwen2_moe
  测试指令：

  cd llm
  # block attn
  python ./predict/predictor.py --model_name_or_path Qwen/Qwen1.5-MoE-A2.7B --dtype float16 --mode dynamic --decode_strategy greedy_search --inference_model 1 --block_attn 1 --append_attn 0
  
  # append attn
  python ./predict/predictor.py --model_name_or_path Qwen/Qwen1.5-MoE-A2.7B --dtype float16 --mode dynamic --decode_strategy greedy_search --inference_model 1 --block_attn 0 --append_attn 1
  
  # throw exception
  python ./predict/predictor.py --model_name_or_path Qwen/Qwen1.5-MoE-A2.7B --dtype float16 --mode dynamic --decode_strategy greedy_search --inference_model 1 --block_attn 0 --append_attn 0

  存在问题：
  block attn 跑不通 在实现AutoModel的时候还是跑的通的（develop也是）
  
  append attn跑不通（develop 也是）
  

[paddle-bot]

Thanks for your contribution!

[codecov]

Codecov Report

Attention: Patch coverage is 0% with 47 lines in your changes missing coverage. Please review.

Project coverage is 52.67%. Comparing base (fb60645) to head (77996bd).
Report is 10 commits behind head on develop.

Files with missing lines	Patch %	Lines
...dlenlp/experimental/transformers/llama/modeling.py	0.00%	15 Missing ⚠️
...dlenlp/experimental/transformers/bloom/modeling.py	0.00%	11 Missing ⚠️
...lp/experimental/transformers/qwen2_moe/modeling.py	0.00%	6 Missing ⚠️
...dlenlp/experimental/transformers/qwen2/modeling.py	0.00%	5 Missing ⚠️
...p/experimental/transformers/chatglm_v2/modeling.py	0.00%	4 Missing ⚠️
...enlp/experimental/transformers/mixtral/modeling.py	0.00%	4 Missing ⚠️
paddlenlp/transformers/auto/modeling.py	0.00%	2 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #9770      +/-   ##
===========================================
- Coverage    52.70%   52.67%   -0.03%     
===========================================
  Files          731      730       -1     
  Lines       117313   114591    -2722     
===========================================
- Hits         61827    60365    -1462     
+ Misses       55486    54226    -1260     

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