Loading activity…
@classiclarryd
Technical Staff at Open Athena, working on Marin
AI
A post says custom kernels let Marin expand the model from 360 billion to 535 billion parameters while speeding up token processing.
AI
A user called a linked post a “huge breakthrough” but did not say what happened.
AI
A post credits Jan Varho with a 0.4-second improvement using a simple idea: masking token continuations that cannot occur in the tokenized dataset.
AI
A post reports a 5.8-second improvement, with changes including a smaller query/key dimension, fused Triton operations and more FP8 computation.
