Skip to content

test_gpu fails on RTX 3090 with CUDA 11.1  #89

@mdabek

Description

@mdabek

Thank you for the warp-transducer, I was successfully using it with previous CUDA versions and on older card (RTX 2080) and with previous Ubuntu 18.04. I am currently trying to change the setup and move to the following configuration:
HW: RTX 3090Ti
OS: Ubuntu 20.04
CUDA: 11.1

The first change I needed to make was to remove the compute_30, since it is not supported by Cuda 11.1. The code can be compiled and I am able to run test_gpu. However, the option_test in test_gpu fails.
Here is the output:

Running gpu tests
finish small_test 0
grad mismatch in options_test expected grad: -0.186844 calculated score: 0 !(-0.186944 < 0 < -0.186744)
grad mismatch in options_test expected grad: -0.062555 calculated score: 0 !(-0.062655 < 0 < -0.062455)
grad mismatch in options_test expected grad: 0.249399 calculated score: 0 !(0.249299 < 0 < 0.249499)
grad mismatch in options_test expected grad: -0.203377 calculated score: 0 !(-0.203477 < 0 < -0.203277)
grad mismatch in options_test expected grad: 0.202399 calculated score: 0 !(0.202299 < 0 < 0.202499)
grad mismatch in options_test expected grad: 0.000977 calculated score: 0 !(0.000877 < 0 < 0.001077)
grad mismatch in options_test expected grad: -0.141016 calculated score: 0 !(-0.141116 < 0 < -0.140916)
grad mismatch in options_test expected grad: 0.079123 calculated score: 0 !(0.079023 < 0 < 0.079223)
grad mismatch in options_test expected grad: 0.061893 calculated score: 0 !(0.061793 < 0 < 0.061993)
grad mismatch in options_test expected grad: -0.011552 calculated score: 0 !(-0.011652 < 0 < -0.011452)
grad mismatch in options_test expected grad: -0.08128 calculated score: 0 !(-0.08138 < 0 < -0.08118)
grad mismatch in options_test expected grad: 0.092832 calculated score: 0 !(0.092732 < 0 < 0.092932)
grad mismatch in options_test expected grad: -0.154257 calculated score: 0 !(-0.154357 < 0 < -0.154157)
grad mismatch in options_test expected grad: 0.229433 calculated score: 0 !(0.229333 < 0 < 0.229533)
grad mismatch in options_test expected grad: -0.075176 calculated score: 0 !(-0.075276 < 0 < -0.075076)
....
score mismatch in options_test expected score: 4.28065 calculated score: -0 !(4.28055 < -0 < 4.28075)
score mismatch in options_test expected score: 3.93844 calculated score: -0 !(3.93834 < -0 < 3.93854)
finish options_test 0
finish inf_test 0
finished 0
Some or all tests fail

All the grads calculated by option_test are 0.
Is this known issue? How can it be debugged and solved?

Thank you!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions