← all posts site ↗

directml can't train my audio models

every time i try to train a small audio model on the windows desktop, the gpu path dies in the same place. torch's stft and istft raise an invalid data type error on directml, which is a problem because the whole augmentation pipeline, room impulse response convolution, pitch shifting, colored noise, routes through complex stft. directml is in maintenance mode, nobody is fixing it. so the nvidia style story of install the runtime and go does not exist on this card.

the funny part is how little it matters. the actual model head is tiny, the full training run is under a minute of cpu time. the augmentation stage is roughly forty times the compute of training, and it is plain audio dsp, so the cpu sets the pace either way. a working gpu saves about two minutes. this is the case where the gpu is the wrong tool because the job is too small for it.

rocm on windows is also no longer the dead end most people still believe. amd's hip sdk lists the rdna4 card as supported on windows and the community wheel index has torch wheels for it. one of my research agents called it linux only from old notes, the amd docs and a live wheel index disagreed. so there is a real gpu path forward if i ever care, but for a model this size, cpu threads and no graphics stack is the correct answer.