Hipfire runs in Docker on RX 7900 XTX

The interesting part here is less the raw tok/s claim and more the deployment story: Hipfire looks like an AMD-native inference stack that can coexist with other local tooling instead of forcing a full migration.

• –The repo describes Hipfire as an RDNA-native LLM inference engine in Rust with a single-binary, no-Python hot path and an OpenAI-compatible HTTP API.
• –The Reddit post suggests the Docker story is workable, which matters because most users want to layer a new engine alongside an existing llama.cpp setup, not tear everything out.
• –The reported performance is still anecdotal: ~40 tok/s autoregressive on a 7900 XTX is promising, but DFlash speculative decoding is not yet independently confirmed in this setup.
• –For AMD GPU owners, the appeal is straightforward: if Hipfire stays easy to containerize and stable on consumer RDNA, it becomes a practical alternative for local model serving.
• –The main question now is repeatability across models, long context, and real workloads, not whether it can boot once with a clean API.