/test full

FYI it seems the 13.0 docker images are up!

FYI it seems the 13.0 docker images are up!

I guess our CI is one of the very first user 😄

I discussed with @leofang offline and found that we need to adjust CUDA header discovery logic to support the change in PyPI CUDA packages. I'm looking into it.
https://developer.nvidia.com/blog/whats-new-and-important-in-cuda-toolkit-13-0/#wheel_package_changes_to_cuda_130

edit: done in 3ad708a

6e7dabe:

• https://ci.preferred.jp/cupy.linux.cuda130/197967/

Seems number of tests are failing only on Linux. (full log is here: https://storage.googleapis.com/chainer-artifacts-pfn-public-ci/cupy-ci/197967/log.txt)

Tests are failing because the fft static library doesn't support sm_75, I will push a fix

$ cuobjdump --list-elf /usr/local/cuda-13/lib64/libcufft_static.a

ELF file    1: libcufft_static.1.sm_80.cubin
ELF file    2: libcufft_static.2.sm_90.cubin
ELF file    3: libcufft_static.3.sm_100.cubin
ELF file    4: libcufft_static.4.sm_110.cubin
ELF file    5: libcufft_static.5.sm_103.cubin
ELF file    6: libcufft_static.6.sm_120.cubin

Seems the flexCI instances has CC 7.5 GPUs so we need to skip these tests :D

FAILED cupy_tests/core_tests/test_raw.py::TestRaw_param_3_{backend='nvcc', in_memory=False}::test_dynamical_parallelism
FAILED cupy_tests/core_tests/test_raw.py::TestRawPicklable_param_0_{compile=False, raw='ker'}::test_raw_picklable
FAILED cupy_tests/core_tests/test_raw.py::TestRawPicklable_param_1_{compile=False, raw='mod'}::test_raw_picklable
FAILED cupy_tests/core_tests/test_raw.py::TestRawPicklable_param_2_{compile=False, raw='mod_ker'}::test_raw_picklable
FAILED cupy_tests/core_tests/test_raw.py::TestRawPicklable_param_3_{compile=True, raw='ker'}::test_raw_picklable
FAILED cupy_tests/core_tests/test_raw.py::TestRawPicklable_param_4_{compile=True, raw='mod'}::test_raw_picklable
FAILED cupy_tests/core_tests/test_raw.py::TestRawPicklable_param_5_{compile=True, raw='mod_ker'}::test_raw_picklable
FAILED cupy_tests/cuda_tests/test_compiler.py::TestNvrtcArch::test_compile_cuda9

test_compile_cuda9 seems to be already fixed, the RawKernel tests work fine in my local machine with cc 7.5, may be related to flexCI settings?

I've re-kicked CUDA 13 CI with the latest commit. Let's see how it goes...
#9289 (comment)

https://ci.preferred.jp/cupy.linux.cuda130/197990/ is the CI results for the commit 8d91920.

In addition to Test1dCallbacks, TestNdCallbacks needs to be skipped on sm_75 as well. I've pushed a fix for that.

For test_raw.py test failures, I could reproduce locally using NVCC="ccache nvcc". The problem is described in #4421 (comment) but I'm not sure why this started to happen in CUDA 13 🤔
Let's kick other CIs for older CUDA versions to isolate the problem.

/test mini

For test_raw.py test failures, I could reproduce locally using NVCC="ccache nvcc". The problem is described in #4421 (comment) but I'm not sure why this started to happen in CUDA 13 🤔

Not sure what happened... Looks like an I/O error, could it be out of space?

            if code_type == 'ptx':
>               with open(result_path, 'rb') as ptx_file:
                     ^^^^^^^^^^^^^^^^^^^^^^^
E               FileNotFoundError: [Errno 2] No such file or directory: '/tmp/tmpzktuvmkn/preprocess.ptx'

No, this is a longstanding bug in ccache that cannot handle --ptx -dc preprocess.cu or --ptx --device-c preprocess.cu option. For some reason ccache automatically appends -o preprocess.o when proxying to nvcc, where it should not (or should append -o preprocess.ptx). I'm still unsure why this started happen here though... Anyway I've added a workaround.

/test mini

Kicked CUDA 13 CIs again.

https://ci.preferred.jp/cupy.linux.cuda130/198024/
https://ci.preferred.jp/cupy.linux.cuda130.multi/198026/
https://ci.preferred.jp/cupy.win.cuda130/198025/

-> all passed!

@kmaehashi Failed to backport automatically.

Not cleanly cherry-picked