So, what are you going to do with these findings? Will you need to get Apple involved? Will you submit pull requests to the torch or mlx folks? Any chance of further improvements with more processing time? Can the regressions be mitigated with some kind of switch between the standard kernels and your optimized ones?
Aaron Reitz
Adreitz
·
AI & ML interests
None yet
Recent Activity
new activity 10 days ago
LAXMAYDAY/FLUX.2-dev-int8-tensorwise:Benchmark vs Q8 GGUF? new activity 23 days ago
Qwen/Qwen3.8-27B:We Cracked Qwen3.8-27B Quant: 27GB INT4 that actually thinks (Heretic Edition)Organizations
None yet
Benchmark vs Q8 GGUF?
#1 opened 10 days ago
by Adreitz
This model is completely unusable in ComfyUI. Are you kidding me? It's ridiculous.
2
#2 opened 11 days ago
by 
liuxingsyue
We Cracked Qwen3.8-27B Quant: 27GB INT4 that actually thinks (Heretic Edition)
🔥 2
8
#63 opened 23 days ago
by 
goldhub
unet unexpected: ['model_sampling.sigmas']
#7 opened 9 months ago
by Adreitz
commented on Automated Discovery of High-Performance GPU Kernels with OpenEvolve about 1 year ago
No way on earth to get "an albino woman with white skin and dark hair"
3
#32 opened over 1 year ago
by 
roberto-milazzi
Is it possible to use locally on Mac?
1
#20 opened over 1 year ago
by 
Gikal999
Vae
1
#6 opened almost 2 years ago
by 
MonikaDreamer
liked a
model about 2 years ago
Lowres image synthesis is flawed
3
#23 opened about 3 years ago
by 
Pupochek
Doesn't work well with DreamFusion, generates bad quality images under high guidance
1
#51 opened about 3 years ago
by 
JunzheJosephZhu
What is full form of "IF" or what does it mean in "DeepFloyd IF"?
1
#357 opened over 3 years ago
by 
HimSinghvi
I think I confused IF, so how do I avoid this?
2
#113 opened over 3 years ago
by 
EliAnon