Machine learning, LLMs, and intelligent products.
running some benchmarks on internal fine tunes. vllm is great for throughput but the memory overhead for pagedattention hits hard on smaller gpus. thinking about switching to a pure triton kernel implementation to save vram. anyone actually shipped this or are we just chasing mic…
See more →
we rewrote the AI & ML layer twice this year and the bug is still the same one from march. the new abstraction mostly renamed the function.
Test only
Lando has always said that the McLaren’s have been tricky to drive (that’s why Ricciardo struggled so much), I think in the extended version of this quote he said that he’s never had a McLaren that suited his driving style. He struggled early last year and while the suspension up…
See more →
test only
