ROCm

I tried three engines to run DeepSeek-V4.1-Flash, a 552-billion-parameter MoE, on two 16 GB AMD MI50s. Porting Project Maya's engine to gfx906 took it from 7 to about 10 tokens/s, and a ten-line fix to how it reads the Engram table removed the worst decode freezes. All five of my coding-agent tasks now pass on it.

Two bugs made DFlash2 speculative decoding useless on AMD MI50. I fixed both, added custom HIP kernels that reach 74 tokens/s, got a 200k-token context for coding agents, and wrote two attention kernels that make that context fast. All of it on two 2018 GPUs at 90 W.