Fix train dtype for native_dropout | fix(torchlib)

[justinchuby]

Previously the train attribute in native_dropout was not casted to BOOL. We need to do that because the underlying attribute type for python bool is INT64 when promoted to input. This change also added a test that will catch the old error.

[codecov]

Codecov Report

Merging #869 (a56db2d) into main (1fc87c3) will increase coverage by 0.02%.
The diff coverage is 84.61%.

@@            Coverage Diff             @@
##             main     #869      +/-   ##
==========================================
+ Coverage   76.55%   76.57%   +0.02%     
==========================================
  Files         112      112              
  Lines       13394    13407      +13     
  Branches     1347     1350       +3     
==========================================
+ Hits        10254    10267      +13     
+ Misses       2806     2805       -1     
- Partials      334      335       +1     

Impacted Files	Coverage Δ
...ipt/tests/function_libs/torch_lib/ops_test_data.py	96.72% <ø> (ø)
...ript/tests/function_libs/torch_lib/extra_opinfo.py	97.34% <83.33%> (-0.96%)	⬇️
onnxscript/function_libs/torch_lib/ops/core.py	77.00% <100.00%> (+0.09%)	⬆️

[titaiwangms]

Is it like the difference between using it as an attribute or an input? Looks like it's going to be used as input of op.Dropout, so we need to promote it to BOOL.

[BowenBao]

This comment is for 1 line above, train: bool = True

I encountered another op with this recently too. It seems train is created as Constant w/ INT type. Is it possible to fix the root to have it create as BOOL type Constant?

[justinchuby]

I think we will need to update the translator. Created #872

[justinchuby]

Is it like the difference between using it as an attribute or an input? Looks like it's going to be used as input of op.Dropout, so we need to promote it to BOOL.

That's correct.