Download docs/LINEAGE.md from staccs/ds4-winhip: direct link, hf CLI and curl.
- Browser
- Download file 6.41 kB
-
https://huggingface.co/staccs/ds4-winhip/resolve/main/docs/LINEAGE.md
- Command line
-
hf download hf://staccs/ds4-winhip/docs/LINEAGE.md
-
curl -L -o LINEAGE.md https://huggingface.co/staccs/ds4-winhip/resolve/main/docs/LINEAGE.md
Project lineage
ds4-dfm-rs is an independent repository, not an independent invention. It
continues the same Git history and MIT-licensed code line:
antirez/ds4
โ CUDA and batched serving
Entrpi/ds4
โ DFM model families and DGX Spark specialization
Baekpica/ds4
โ Rust host promotion
Baekpica/ds4-dfm-rs
No history filter, squash, or clean-room relabel was used for the split. Git
authors and ancestry remain intact, including the original fork point
e16ead1e29c81a67bbb64e5b001117679cf9ce6e.
Contributions carried forward
antirez/ds4established DwarfStar: a self-contained, deliberately narrow native engine built around a few useful model and hardware combinations. It also established the project's explicit debt to llama.cpp and GGML.Entrpi/ds4developed the CUDA/Linux and batched multi-request serving line used here, with DGX Spark / GB10 as a reference target.Baekpica/ds4added the DFM line: explicit Solar Open2, K-EXAONE, Motif-3, dots3-note, and Qwen3.8 family paths alongside DeepSeek, plus their persistent state, serving, and CUDA work.Baekpica/ds4-dfm-rscontinues that line with the host and control plane promoted to Rust over the same narrow native backend.
The Rust split changes ownership boundaries and the release lifecycle. It does not erase the native backend, change the origin of inherited code, or claim that ds4 was rewritten from scratch.
Immutable split references
| Ref | Commit | Meaning |
|---|---|---|
v0.6.5-dfm |
d02e2a4777a34a9f52fd987453b3ea1801fac52e |
Entrpi v0.6.5 reconciled into the DFM line; immutable C baseline. |
| Qwen freeze | 4d40d97f1e575400237a6e5cef21d7f74404a38d |
Last frozen Qwen C behavior used by the resumed Rust campaign. |
| Rust implementation cut | d126e56877390e9522dde333a34f0d582c3e246c |
Promoted Rust-host implementation after the final server fix. |
ds4-dfm-rs-genesis |
fe7733fb4f7e18204b6ea0a00fe3b136d2029b17 |
Annotated split point after SPLIT_READINESS.md became green. |
| Post-genesis fix | 7ae5257de96e7bae807e6c406e45e17ba347e52f |
Preserves the closed-hop error through a distributed enqueue race. |
The original destination repository contained a small rename scaffold at
b01d1fa4172a5c957fe1232774629a192493efe4. It remains recoverable through the
annotated pre-genesis-scaffold-b01d1fa tag; the default branch now follows the
continuous DwarfStar ancestry above.
The annotated v0.1.0-rc.1 tag names the first independent Rust-host release
candidate containing this document.
v0.1.0-rc.2 retains the same runtime behavior and sequences two mutating C
oracle calls explicitly so host parity is deterministic across x86 and ARM C
compilers.
v0.1.0-rc.3 continues the same native backend and model-family lineage while
hardening the shared Rust serving plane for agent concurrency, streaming,
continuations, admission, cache accounting, and shutdown. Its Qwen Q5+Sidecar
one- and two-bank runs validate those host changes; they do not redefine the
v0.6.5-dfm golden or claim a new model/kernel release.
v0.1.0-rc.4 selectively ports the GLM 5.3 Flash graph and vision tower from
official antirez/ds4 commit
110afdd8886586f18fc9b28bc5533152dd10e728 into the existing Rust/native
boundary. Its support claim is limited to the exact Q2 main GGUF and vision
sidecar validated on DGX Spark; it does not import a second host stack or claim
the wider upstream GLM backend matrix.
Versioning after the split
v0.1.0 is the first independent Rust-host baseline with
stable host/runtime ownership and native performance observability through
ds4-perf. Its ledger records production qualification and workload limits.
The native lineage remains unchanged.
v0.6.5-dfm remains the behavioral baseline, not the new repository's version.
ds4-dfm-rs starts at v0.1.0-rc.1 so Rust-host releases can advance without
implying a new Entrpi or antirez release. Old tags remain available as
provenance and comparison points.
Upstream port policy
The repositories no longer have an automatic merge relationship. Changes from antirez, Entrpi, or Baekpica ds4 are reviewed as selective semantic ports:
- record the source repository and exact commit;
- preserve original authorship and license notices;
- port only the behavior needed by a currently supported path;
- adapt it at the Rust/native ownership boundary without importing a parallel host stack;
- rerun the affected C/Rust parity, model, and performance gates.
Likewise, a generally useful fix developed here should stay small and close to the original native style so it can be proposed upstream without dragging in the Rust host or DFM-only policy.
Development context
DwarfStar follows useful open weights that fit personal and workstation-class machines. Routed-expert quantization, compressed or recurrent state and fast local SSDs make large models and long contexts practical. Support remains opportunistic and specific to useful model/artifact/hardware combinations; a family may be retired when a better model replaces it.
The Rust continuation keeps the same approach. DFM (๋ ์ ํ์ด๋ฐ์ด์ ๋ชจ๋ธ, ๋ ํ๋ชจ) and adjacent families have distinct tensor layouts, state machines, prompt protocols and kernels. Rust makes lifecycle and serving policy easier to own while the native engine stays close to the hardware. Family extensions start from a validated path and rerun its correctness and performance gates.
Development uses strong coding-agent assistance. Humans lead ideas, scope, testing and debugging, and own the resulting measurements and behavior. The largely hand-built work in llama.cpp/GGML, the original DwarfStar and Entrpi's CUDA-serving line remains the foundation; assistance does not change that provenance.
Third-party code
The repository remains MIT licensed; see ../LICENSE. Existing
copyright notices are preserved. Selected quantization definitions, CPU paths,
and native kernels derive from or were adapted from
llama.cpp and GGML under the MIT
license. The MMQ pin and local patch policy are recorded in
../cuda/mmq/VENDOR.md.