Awni Hannun
Awni Hannun
The next step is to remove the Python one and make a binding in python/src/fft.cp
This PR is missing tests.
How long are the files that cause issues? It looks like a 1GB of audio? What's the sample rate / duration?
You can do `ffmpeg -i 1f7afc6c-246e-4d6c-943b-0223cc4f27e5_full.m4a` and share it here?
I think the problem is computing the log mel spectrograms is pretty memory intensive when the audio gets long. It should be doable to batch that computation.
I think it's a good idea to report the error as you point out. Regarding allowing CPU on the sim -- generally MLX workloads are quite compute intensive and it's...
This seems to be related to graph exec update. If you disable cuda graphs the tests run fine `MLX_USE_CUDA_GRAPHS=0`. Also if you use a fresh `graphExec` for each graph the...
A standalone C++ script that reproduces the error. Instead of getting the desired output, on B200, I get: ``` y[0..3] = {3.7793, 3.79102, 3.80078, 3.81055} Graph exec update success! y[0..3]...
Closed in #2813
Thanks! We don't have concrete plans for this yet. Some activities in this direction are underway. CC @nastya236 you could chat with about it.