Open
Description
🐛 Describe the bug
This is exposed by #145904
Test for fft_hfftn fails a size/stride assertion. This most likely means the op has inconsistent meta v.s. eager implementation regarding the output strides.
Repro:
TORCHINDUCTOR_SIZE_ASSERTS=1 PYTORCH_OPINFO_SAMPLE_INPUT_INDEX=0 python test/inductor/test_torchinductor_opinfo.py TestInductorOpInfoCUDA.test_comprehensive_fft_hfftn_cuda_float16
Similar failures happens for
- fft_fft (https://github.com/pytorch/pytorch/actions/runs/13023114885/job/36328152122 )
- fft_ifft
- fft_ihfft
- fft_ihfft2
- fft_rfft
- fft_rfft2
But I could not repro those locally.
Error logs
No response
Versions
.
cc @chauhang @penguinwu @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @amjames @desertfire @aakhundov