DwarfStar (antirez/ds4) as a backend: 1bit serve --device ds4

[bong-water-water-bong]

DwarfStar (MIT) is a native engine for a few large MoE models (DeepSeek V4/V4.1 Flash, GLM 5.x Flash, Qwen3.8-Flash-Next) in its own GGUF layouts, with first-class Strix Halo ROCm support and SSD expert streaming. Its GGUFs are the most downloaded DeepSeek V4.1 Flash files on HF (antirez/deepseek-v4.1-flash-gguf, 940k), and llama.cpp cannot load them. This adds it as a child backend, like ZINC.

• third_party/ds4: antirez/ds4 0aaea5a238fb; bump-ds4.yml follows upstream main daily.
• scripts/build-ds4.sh: builds a copy of the tree (the submodule stays clean) for rocm / cuda / metal / cpu. ROCm finds TheRock's SDK and passes its include/ with -isystem (clang otherwise searches it after /usr/include, where a distro HIP of another version shadows it). DS4_TEST=1 runs DwarfStar's model-free routed-MoE GPU test.
• -DONEBIT_DS4=ON; 1bit serve --device ds4 with --ctx-size, --ssd-streaming, --ds4 PATH. A child now carries its own readiness path: ds4-server has no /health and opens its port once the model is loaded, so serve waits on /v1/models.
• docs/dwarfstar.md; serve.md, README (repo XDNA: pin upstream amd/xdna-driver (and its XRT), build it privately, keep it current #12) and NOTICE.

Verified on Strix Halo (TheRock ROCm):

Check	Result
DS4_TEST=1 scripts/build-ds4.sh build/ds4 rocm	builds; routed-MoE MXFP4 test PASS (0 failures; variants bitwise OK); submodule clean
engine -DONEBIT_DS4=ON	builds; ctest 7/7
serve --device ds4 -m <missing>	DwarfStar's "cannot open model", then serve exits (backend did not become ready)
--ssd-streaming on another device	refused

Not yet run: a full model. DeepSeek V4 Flash Q2 is 81 GiB on disk and needs about that much free memory; the box has 89 GB of disk and ~60 GB of memory free while other work runs. Hold the merge until it answers a chat through serve --device ds4.

🤖 Generated with Claude Code

[context7]

Docs7 for 1bit-monster/engine

Result	Status	Action
Deployment	➖ Not used	—
Content review	➖ Did not run. This site has no agent runs available this month. Wait for the monthly reset or check your Docs7 plan.	—

Commit 1a9ca1d

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 12 - Partially compliant Compliant requirements: Pin the XDNA userspace stack to upstream amd/xdna-driver with commit 5d302c9 Build XRT and the XDNA shim plugin privately under a prefix Exclude XRT's OpenCL layer (xocl) from the build Use the build with -DONEBIT_XRT_ROOT=/root for the NPU lane Keep the pin current with a daily job that opens a PR when upstream moves Update docs/npu.md with a section on "The XDNA stack" Non-compliant requirements: None Requires further human verification: None

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Incorrect ready path for ds4 The ds4 device uses /v1/models as its ready path, but this path may not be consistently available or correctly indicate model readiness across different versions of the ds4-server. This could lead to incorrect startup behavior or premature readiness signaling. l.ready_path = "/v1/models"; Missing validation for ds4 model file The code does not validate that the model file specified for the ds4 device actually exists or is a valid DwarfStar GGUF file before attempting to launch the ds4-server. This could lead to runtime errors or unexpected behavior if an invalid model is specified. argv = {o.ds4.empty() ? default_ds4() : o.ds4, "-m", o.model,