Max Laves

Sr. research scientist @ImFusionGmbH | Prev. @PhilipsResearch, @mtec_tuhh, @UniHannover @[email protected]

[Max Laves @MaxLaves]

@luciascarlet My iPhone 15 sometimes loses 30% battery over night while doing nothing (80% to 50% in 8 hours). Battery usage screen shows nothing?!

[Max Laves @MaxLaves]

@luciascarlet There‘s a recent piece by Michael MJD on exactly this https://t.co/jFP7Ee5IxH

[Andre Klekamp 🇪🇺 @KlekampAndre]

Wir haben eine Petition gestartet!

Für weitere Investitionen in #Osnabrücker #Bildung, mehr Unterstützung durch den Bund und gegen Einsparungspläne der CDU!

Bitte teilen und supporten! 🙏🏻

Unterschreiben: #Initiative Osnabrücker Bildungsappell:

https://t.co/rQ1sWRIECl

[Max Laves @MaxLaves]

@vbot30000 @__tinygrad__ Both are using Metal. It comes down to different implementations of conv op (gemm conv vs. winograd etc).

[Max Laves @MaxLaves]

@unixpickle That does indeed sound like a bug in MPS.

[Max Laves @MaxLaves]

@unixpickle Interesting! I‘ve trained 2D conv nets using MPSGraph and do not remember convergence issues.

[Max Laves @MaxLaves]

@unixpickle Are you using the gradients as provided by MPS methods or did you implement it yourself?

[Max Laves @MaxLaves]

@tepienie @luciascarlet +1 for VAG Rounded!

[Max Laves @MaxLaves]

@bhdebayan @AravSrinivas Nice visuals!

[Max Laves @MaxLaves]

@RamanDutt4 Matplotlib + seaborn for graphs, Affinity Designer for everything else.

[Max Laves @MaxLaves]

@unixpickle @vgoklani_ai @awnihannun Sad. I wonder why Apple won’t let us use it.

[Max Laves @MaxLaves]

@unixpickle @vgoklani_ai @awnihannun There‘s no API for using the ANE for anything but CoreML models, right?

[Max Laves @MaxLaves]

@unixpickle Ah, good point. I think I was only benchmarking the forward pass.

[Max Laves @MaxLaves]

@marksaroufim @soumithchintala Better Apple silicon support.

[Max Laves @MaxLaves]

@ivanfioravanti Wouldn’t an open lid help with heat dissipation? My MBP get‘s pretty toasty in the area between keyboard and screen.

[Max Laves @MaxLaves]

@matifanger @_akhaliq https://t.co/G24NbBx3A1

[Max Laves @MaxLaves]

@ivanfioravanti @LMStudioAI Ah nevermind, I accidentally downloaded the non-instruct model. The 32B 4bit instruct model works fine with MLX in LM Studio.

[Max Laves @MaxLaves]

@KassinosS @awnihannun Do you see a performance cliff for large input tensors to your 3D convolution? I.e., performance benefits shrinking the larger the input tensors are?