jad-nohra

4 ms·

1. 1.▲
   by jad-nohra 8mo ago

   Thank you I will take a look at that. It looks like MoE models like Llama-4-Scout report active params (17B) not total (~109B). Some MoE models already work correctly when HuggingFace has the full metadata. Fix is in progress. Fixed in v0.3
2. 2.▲
   Show HN: Vram.run – Compare API providers, local GPUs, and cloud for any model(vram.run)

   1 points by jad-nohra 8mo ago|2 comments
3. 3.▲
   by jad-nohra 9mo ago

   I've been working on a Rust course that takes a different approach from what I could find. I can't memorize rules. God knows I've tried. But if I understand the system that produces them I can, because systems have structure