Lemonade v11.8.0: the 15-backend SDK lands ds4 for Strix Halo
AMD shipped the 15-backend Lemonade SDK — and the new one is literally built for our hardware. DwarfStar (ds4) runs DeepSeek V4 Flash on Strix Halo's gfx1151, and the whole release is now compiled into the same binary as our engine.
Milestone: the AMD SDK release built for our chip — the 15th backend (DeepSeek V4 Flash for Strix Halo) embedded, plus download resume and rate limits. Full technical write-up → Lemonade compat guide
The release that takes the SDK to 15 backends
On August 26 AMD shipped Lemonade v11.8.0 — the release that pushes the SDK to 15 backends — and the 1bit.MONSTER engine re-vendored its embedded server core the same day. Because we embed Lemonade's server core as a library (compiled into build/1bit, no subprocess, no wrappers), every feature in the release arrived in our binary the moment the re-vendor landed: all 15 backends, the router, the full API surface, and every breaking change.
It was also a notably personal release for this project: the changelog credits @bong-water-water-bong — our PR #2462, cancelling model downloads on Ctrl+C via a libcurl progress callback, shipped upstream and now runs inside our own embed.
ds4: DwarfStar, DeepSeek V4 Flash, Strix Halo only
The headline backend is ds4 (DwarfStar4) — antirez's self-contained DeepSeek V4 inference engine with an OpenAI-compatible ds4-server. The Lemonade catalog entry is DeepSeek-V4-Flash-IQ2XXS-DS4 (chat, reasoning, tool-calling), and the support row is refreshingly specific: Linux, ROCm, gfx1151 only — "Prebuilt ds4 for AMD Strix Halo".
That is exactly the silicon our inference box runs. The build is pinned (b0001) from lemonade-sdk/ds4-rocm, which compiles a pinned upstream commit and bundles the ROCm runtime alongside the binary — so ds4 needs no system ROCm install, and it stays out of the way on machines without a Strix Halo GPU. Experimental for now, chat-mode only, but it is compiled into our server core and ready to surface on strixhalo.
What else landed
New catalog models: RPG-HaloTales-V2, Flux-2-Klein-4B/9B, and new OpenMOSS voice-design and sound-effect models — all verified live in our 124-model registry. Cloud providers gained custom auth header names/prefixes and byte-for-byte Anthropic Messages passthrough via wire_format. Downloads can be cancelled with Ctrl+C and resumed, capped by a new download_rate_limit config key, with TCP keepalive and streaming heartbeats keeping long connections alive. A lemonade update-models command plus global and per-model auto-update settings keep installed models current.
The breaking changes are all server-side and all in our binary: --port/--host are now ephemeral overrides, custom *_args precedence was reworked, llama.cpp defaults to --parallel 1, telemetry attributes moved to the standard session.id/user.id keys, whisper transcription defaults to auto-detected language, and the undocumented /v1/params routes are gone.
The re-vendor, repeatable
This was the second run of the loop we formalized with v11.7.0: pull the pinned commit, re-apply the embeddability patch, fix any engine-facing API drift, build, smoke-test. v11.8.0 changed the Server constructor (new config_dir parameter) — both of our embed call sites were updated in the same commit. The local patch (CMAKE_CURRENT_SOURCE_DIR, system json/httplib reuse, PUBLIC include dirs, guarded test police) applied clean with no new guards needed.
One behavior change worth knowing: /v1/models now hides backends with no runnable models on the host — the full registry is behind ?show_all=true. On a non-Strix-Halo machine ds4 quietly filters itself out; on strixhalo it appears automatically.
15 backends in one ELF, the 15th built for the exact GPU we own. The SDK sync loop keeps it that way after every upstream release.
Run it
build/1bit unified --lemonade --port 8088 — the same binary serves all 15 backends through the OpenAI-compatible API. On a Strix Halo box, DeepSeek-V4-Flash-IQ2XXS-DS4 is one /v1/chat/completions call away, ROCm runtime bundled, no system install.