Inference stacks, on the wrong hardware.
The world opens itself before those with noble hearts.
Embedded background — Verilog, ESP32, RISC-V. Mostly kernels and inference now: finding where an abstraction assumed hardware it doesn't have, proving the replacement is numerically identical, then making it fast.
Currently getting DeepSeek-V4 to serve on SM80. Upstream at sgl-project/sglang.


