01 / Llama 2 decode / RX 6700 XT
13.93% Faster Decode on an AMD RX 6700 XT.
A Vulkan optimization for Llama 2 7B Q4_0 on one AMD RX 6700 XT.
- decode throughput
- +13.93%
- tokens/s
- 85.71
- matched protocol
- ABBA
~/projects
A few projects I think you'll find cool.
01 / Llama 2 decode / RX 6700 XT
A Vulkan optimization for Llama 2 7B Q4_0 on one AMD RX 6700 XT.
02 / Coup RL bot
An LSTM actor-critic trained with PPO against a league of its past selves to bluff, challenge, block, and guess your hand in 1v1 Coup. The engine and the network run entirely in your browser.
03 / Heston regime lab
Heston, Bates, and two-factor Heston models calibrated live in the browser to a real S&P 500 option chain, with market regimes found by a hidden Markov model.