jackxlau··on Don't trust large context windowsIn my own testing I have seen peak performance happen usually within 15-20% of the intended context limit, albeit there are a few optimizations depending on the task quality.
jackxlau··on The MilkV Jupiter 2/SpacemiT K3 (RISC-V vector compute)the handshake process to unlock the A100 cores is pretty interesting it reminds me of driving CPU cores like a GPU command queue.
jackxlau··on Software Is Made Between CommitsI came across the conclusion here since a change sometimes spans several repos, per-repo history optimizes the wrong target.