4.2 KiB
ds4 (DwarfStar) — user guide
A from-source build of antirez/ds4, a local
inference engine for DeepSeek V4 Flash/PRO. This is a short operational
guide; see the upstream README for the full
documentation (server API, agent, distributed inference, tuning, etc.).
What's installed
The package builds for a chosen GPU backend (cpu, rocm, or cuda) and
installs these binaries:
| Binary | What it does |
|---|---|
ds4 |
Interactive CLI chat / one-shot -p "..." |
ds4-server |
OpenAI/Anthropic/Responses-compatible HTTP API |
ds4-agent |
Native terminal coding agent (alpha) |
ds4-bench |
Throughput benchmarking |
ds4-eval |
Capability/regression eval harness |
ds4-download-model |
Helper to fetch GGUF weights (see below) |
There is no graphical UI. ds4-server exposes a URL you can point an
OpenAI-compatible client at.
1. Get the weights
Weights are not packaged. They live at the project's Hugging Face repos
(antirez/deepseek-v4-gguf
for DeepSeek V4, antirez/GLM-5.2-GGUF
and unsloth/GLM-5.2-GGUF for
GLM 5.2) — public, MIT, no token needed — and only these custom GGUFs work, not
arbitrary DeepSeek/GLM GGUFs.
ds4-download-model writes into ./gguf and links ./ds4flash.gguf in the
current directory (override the root with DS4_HOME, or the gguf dir with
DS4_GGUF_DIR):
# DeepSeek V4 Flash (0731 checkpoints)
ds4-download-model ds4f-q2 # ~81 GB, the pick for 96/128 GB machines
ds4-download-model ds4f-q2-q4 # ~98 GB, last 6 expert layers at q4
ds4-download-model ds4f-q4 # ~153 GB, for >=256 GB machines
ds4-download-model ds4f-mxfp4 # ~156 GB, native MXFP4 (Metal/CUDA)
ds4-download-model ds4f-dspark # ~6 GB, optional DSpark speculative draft
# DeepSeek V4 PRO (512 GB / distributed)
ds4-download-model pro-q2-imatrix # single-file PRO q2
ds4-download-model pro-q4-split # both halves of PRO Q4 split
# GLM 5.2
ds4-download-model glm-unsloth-q4 # Unsloth UD-Q4_K_XL, 11 shards
ds4-download-model glm-antirez-iq2xxs # routed IQ2_XXS, reduced-memory
ds4-download-model glm-antirez-q2 # routed Q2_K, ~262 GB
ds4-download-model glm-antirez-q4 # routed Q4_K, ~434 GB
ds4-download-model --help # full target list
Pick by machine memory: ds4f-q2 (96/128 GB), ds4f-q4 (>=256 GB),
pro-* (512 GB / distributed), glm-* (per-target sizes above). The small
Flash quants download via curl; the MXFP4, PRO, and GLM files need the hf
CLI (not bundled — nix shell nixpkgs#huggingface-hub).
2. Run it
# interactive chat (defaults to ./ds4flash.gguf in the current dir)
ds4
# explicit model path
ds4 -m /path/to/gguf/DeepSeek-V4-Flash-IQ2XXS-...-imatrix-0731.gguf
# DSpark speculative decoding (greedy only; needs the ds4f-dspark support GGUF)
ds4 -m ./ds4flash.gguf \
--dspark --mtp ./gguf/DeepSeek-V4-Flash-DSpark-support-0731.gguf --temp 0
# GLM 5.2 (experimental greedy MTP via --glm-mtp)
ds4 -m gguf/GLM-5.2-UD-IQ2_XXS_RoutedIQ2XXS_blk78Q2K.gguf --glm-mtp-timing --temp 0
# HTTP server on the LAN
ds4-server -m ./ds4flash.gguf \
--ctx 100000 --kv-disk-dir ./kv --kv-disk-space-mb 8192 --host 0.0.0.0
Then point a client at http://<host>:8000/v1 (endpoints: /v1/chat/completions,
/v1/messages, /v1/responses, ...). Use --cors for browser clients.
Backend notes
- rocm targets a specific GPU arch (default
gfx1151, AMD Strix Halo). At runtime the user needs/dev/kfd+ render-node access, and large models need the GTT memory kernel params from upstream'sSTRIXHALO.md. On Strix Halo, exportHSA_ENABLE_SDMA=0(SDMA is buggy on its unified memory). - cuda is unfree (CUDA closure) and needs an NVIDIA driver at runtime.
- cpu is a reference/debug path only — not for production inference.
Maintenance
This tracks upstream main (no releases). To bump, set rev in default.nix
and refresh hash. Binaries are built per-host (-march=native), so build each
GPU variant on its target host. See default.nix's header comment for details.