Refine deepseekv2 modeling for to_static

[zhangbo9674]

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

Bug fixes

PR changes

Others

Description

Refine deepseekv2 modeling for to_static

[paddle-bot]

Thanks for your contribution!

[codecov]

Codecov Report

Attention: Patch coverage is 0% with 4 lines in your changes missing coverage. Please review.

Project coverage is 51.34%. Comparing base (5eeb7aa) to head (4b83dee).
Report is 22 commits behind head on develop.

Files with missing lines	Patch %	Lines
paddlenlp/transformers/deepseek_v2/modeling.py	0.00%	3 Missing ⚠️
...addlenlp/transformers/deepseek_v2/modeling_auto.py	0.00%	1 Missing ⚠️

❌ Your patch check has failed because the patch coverage (0.00%) is below the target coverage (80.00%). You can increase the patch coverage or adjust the target coverage.
❌ Your project check has failed because the head coverage (51.34%) is below the target coverage (58.00%). You can increase the head coverage or adjust the target coverage.

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #9851      +/-   ##
===========================================
- Coverage    51.54%   51.34%   -0.21%     
===========================================
  Files          741      745       +4     
  Lines       117904   118584     +680     
===========================================
+ Hits         60775    60886     +111     
- Misses       57129    57698     +569     

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

[DrownFish19]

如果attention_mask是支持use_cache的版本，那么就不是casual_mask，或者推理时使用left-padding，那么attention_mask也不是casual_mask，这里的修改不能覆盖之前的情况

[zhangbo9674]

这里的改动，主要是因为动转静不支持如下场景：

if self.config.use_flash_attention:
    attention_mask = None if is_casual_mask(attention_mask) else attention_mask

当前这样该是基于目前的场景下 use_flash_attention 下 attention_mask 一定为 None，还有一种改法，就是将控制流判断后移到调用的地方，但是目前这种情况下，自动并行的切分推导对控制流的场景支持还存在一些问题，因此为了不影响后续流程，先按照前面的改法实现

if is_casual_mask(attention_mask):
    layer_outputs = decoder_layer(
                hidden_states=hidden_states,
                position_ids=position_ids,
                attention_mask=None,
                output_attentions=output_attentions,
                past_key_value=past_key_value,
                use_cache=use_cache,
                attn_mask_startend_row_indices=attn_mask_startend_row_indices)
else:
    layer_outputs = decoder_layer(
                hidden_states=hidden_states,
                position_ids=position_ids,
                attention_mask=attention_mask,
                output_attentions=output_attentions,
                past_key_value=past_key_value,
                use_cache=use_cache,
                attn_mask_startend_row_indices=attn_mask_startend_row_indices)

[DrownFish19]

LGTM