pt: avoid set_default_dtype in tests

[njzjz]

Why it is bad:
The default dtype in the producution code is still float32. When setting it to float64 during tests, the actual producation behavior may not be properly tested. (for example, if the production code misses dtype for pt.zeros, we are not able to find it in these tests)

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (ab35468) 75.05% compared to head (622ddff) 75.05%.

Additional details and impacted files

@@           Coverage Diff           @@
##            devel    #3303   +/-   ##
=======================================
  Coverage   75.05%   75.05%           
=======================================
  Files         396      396           
  Lines       33895    33895           
  Branches     1604     1604           
=======================================
  Hits        25441    25441           
  Misses       7593     7593           
  Partials      861      861           

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

[wanghan-iapcm]

Why it is bad: The default dtype in the producution code is still float32. When setting it to float64 during tests, the actual producation behavior may not be properly tested. (for example, if the production code misses dtype for pt.zeros, we are not able to find it in these tests)

pt uses float32 only when DP_INTERFACE_PREC is set to low ?

[njzjz]

pt uses float32 only when DP_INTERFACE_PREC is set to low ?

When DP_INTERFACE_PREC is low, GLOBAL_PT_FLOAT_PRECISION is float32. Otherwise, GLOBAL_PT_FLOAT_PRECISION is float64. That's it.

If dtype is not set, a tensor (torch.tensor, torch.ones, torch.zeros, etc) still uses float32, no matter what GLOBAL_PT_FLOAT_PRECISION is.