Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

Support CUDA 10.1 Update 2

See original GitHub issue

Currently CuPy does not work with CUDA 10.1 Update 2. It should be resolved by the final release of v7.

Following tests are failing (
- cupy_tests.fft_tests.test_fft.TestRfftn_param_11.test_irfftn
- tests.cupy_tests.linalg_tests.test_product.TestMatrixPower.test_matrix_power_inv2
#2542 broke compatibility with CUDA 10.1 Update 2. Some undocumented APIs (e.g., cusparseCbsrilu02_bufferSizeExt) are dropped in CUDA 10.1 Update 2. As they are still declared in cusparse.h (this is a bug of cuSPARSE), build succeeds but import cupy fails. (PR #2573)

CUDA 10.1.105-1: ok
CUDA 10.1.168-1: ok
CUDA 10.1.243-1: fail (cuSPARSE 10.3.0.243)

CUDA 10.1 Update 2 now uses semantic versioning, it is better to print version of every library (cusolver, cusparse, …) currently linked to CuPy in cupy.show_config. (PR #2578)
CUDA 10.1 Update 2 changed FP16 headers (PR #2592)

Issue Analytics

State:
Created 4 years ago
Reactions:6
Comments:16 (16 by maintainers)

Top GitHub Comments

1reaction

emcastillocommented, Oct 30, 2019

The API calls that #2573 removed were not present in the stable release 😃

0reactions

leofangcommented, Aug 12, 2020

The issue was reproduced here as well. It seems the issue happens when

Pascal generation GPU is used,

Inverse (complex-to-real) FFT is performed,

Size of array in last transform axis is two, and

CUDA version is 10.1 update 2 or 10.2.

This issue will be fixed in a future CUDA release.

@anaruse Any chance you can check if this bug is fixed in CUDA 11.0 or not?