Got hipfire running in Docker on my RX 7900 XTX alongside llamacpp

Been dealing with long context failures on Qwen3.6 27B and stumbled onto hipfire. Spent an evening dockerizing it so it runs alongside an existing llamacpp stack without touching anything. Running Qwen3.6 27B MQ4 on a 7900 XTX. The TriAttention sidecar and DFlash draft both load correctly per the logs. ~40 tok/s AR, haven't confirmed DFlash is actually engaging yet. Still early but it responds correctly and the API is clean. One thing that tripped me up: hipfire isn't a single binary you just