use tensor.shape bug not paddle.shape(tensor)

[wanghuancoder]

PR types

Others

PR changes

Others

Description

paddle.shape(tensor)在最新的动转静（SOT）中，会造成子图打断，进而影响模型性能。统一改成tensor.shape。

[paddle-bot]

Thanks for your contribution!

[codecov]

Codecov Report

Attention: Patch coverage is 68.63636% with 69 lines in your changes are missing coverage. Please review.

Project coverage is 55.32%. Comparing base (2273ee7) to head (4b044ef).
Report is 20 commits behind head on develop.

Files	Patch %	Lines
paddlenlp/transformers/transformer/modeling.py	0.00%	12 Missing ⚠️
...lenlp/ops/fast_transformer/transformer/decoding.py	0.00%	7 Missing ⚠️
paddlenlp/transformers/bloom/modeling.py	0.00%	5 Missing ⚠️
paddlenlp/transformers/xlm/modeling.py	50.00%	5 Missing ⚠️
paddlenlp/transformers/xlnet/modeling.py	77.27%	5 Missing ⚠️
paddlenlp/generation/utils.py	20.00%	4 Missing ⚠️
paddlenlp/transformers/gpt/modeling_auto.py	0.00%	4 Missing ⚠️
paddlenlp/transformers/layoutxlm/modeling.py	42.85%	4 Missing ⚠️
...s/fast_transformer/transformer/fast_transformer.py	0.00%	3 Missing ⚠️
paddlenlp/trainer/trainer_compress.py	0.00%	2 Missing ⚠️
... and 16 more

Additional details and impacted files

@@ Coverage Diff @@ ## develop #8260 +/- ## =========================================== + Coverage 55.15% 55.32% +0.17%  =========================================== Files 601 614 +13 Lines 91764 95333 +3569 =========================================== + Hits 50611 52747 +2136  - Misses 41153 42586 +1433 

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

[ZHUI]

为什么会出现这个性能问题呢？之前是推荐使用 paddle.shape的，有段时间 tensor.shape 并不能动转静。
后续新增模型是否要监测 padde.shape 的使用。

[ZHUI]

LGTM

[wanghuancoder]

为什么会出现这个性能问题呢？之前是推荐使用 paddle.shape的，有段时间 tensor.shape 并不能动转静。 后续新增模型是否要监测 padde.shape 的使用。

在很早期的基于AST的动转静时，推荐使用paddle.shape，否则无法动转静。但后来AST修复了这个问题。AST也能支持Tensor.shape。
而现在动转静已经默认改用了基于SOT的技术，这种技术对paddle.shape的支持还不够好，会造成动态图转成多个子静态图。paddle.shape在一些场景下会造成子图打断。而静态图子图越多，就会执行效率就会越差。
这个我认为暂时不用监控，因为有些地方是必须使用paddle.shape的。SOT后续也会逐步完善~
但短期，子图打断严重既影响动转静性能，有影响动转静+CINN性能。因此推全修复主要repo~