Conversation
Jarkko Sakkinen
jarkko
Edited 23 days ago
Texture mapping most likely at least inspired tiling and optimizations where representative candidate tokens are cherry picked and combined with a factor clearly seem to do miracles for my Qwen32-VL-2B implementation. Here's benchmark results.
caption, medium 43.3% faster, low 46.2% faster
OCR, medium 44.1% faster, low 44.8% faster
I've now used to more like 0,5-3% increases so this is really motivating bump :-) Aforementioned types of optimizations are also super-intuitive if you've ever worked e.g. with texture mapping or color quantization.
Soon out in the form of new ReadSeek release.
caption, medium 43.3% faster, low 46.2% faster
OCR, medium 44.1% faster, low 44.8% faster
I've now used to more like 0,5-3% increases so this is really motivating bump :-) Aforementioned types of optimizations are also super-intuitive if you've ever worked e.g. with texture mapping or color quantization.
Soon out in the form of new ReadSeek release.
0
0
1