Nano-vLLM: A lightweight vLLM implementation built from scratchgithub.com5 points·by platers·w zeszłym roku·0 comments
platers·7 miesięcy temu·discussthis would pair well with https://gonzoml.substack.com/p/neurips-2025-best-papers-in-c...
platers·11 miesięcy temu·discussI'm struggling to understand where the gains are coming from. What is the intuition for why DiT training was so inefficient?
platers·11 miesięcy temu·discusshttps://suno.com/studio-waitlist Just a waitlist so far, but looks like this is the direction suno is going
platers·w zeszłym roku·discussIt sounds very similar to suno v3.5 (including the audio quality) Likely they trained on suno generations.
platers·w zeszłym roku·discussFlash attention natively supports packing multiple variable length sequences into a single call, what is the advantage of jagged flash attention?