dwarfstar-nix

A Nix flake that builds DwarfStar (ds4) — antirez's DeepSeek-V4 inference runtime — with its ROCm backend for AMD Strix Halo (gfx1151: Ryzen AI MAX, Radeon 8050S/8060S).

ds4 is its own vertical engine (not llama.cpp/ollama), hand-tuned for one model family on one platform. Its upstream build instructions assume Ubuntu + apt; this flake packages the ROCm build for Nix/NixOS so you get the binaries with one command and no manual ROCm setup.

Requirements

• An AMD Strix Halo machine (Ryzen AI MAX APU, Radeon 8050S/8060S = gfx1151) with ≥96 GB unified RAM (128 GB recommended).
• Nix with flakes enabled (experimental-features = nix-command flakes).
• ROCm device access: your user must be in the render group and able to open /dev/kfd:

  sudo usermod -aG render,video "$USER"   # then log out and back in

Quick start

1. Download the model (~81 GB)

DeepSeek-V4-Flash, 2-bit quant — fits 96/128 GB machines. Run this from the directory where you want the model stored; it writes ./gguf/… and symlinks ./ds4flash.gguf.

nix shell github:paolino/dwarfstar-nix -c ds4-download-model q2-imatrix

2. Run a prompt

nix run github:paolino/dwarfstar-nix -- \
  -m ./ds4flash.gguf --rocm --ssd-streaming -p "Hello, who are you?"

Add --nothink for terse answers; omit it for full reasoning.

3. Or run the server (OpenAI-style API)

nix shell github:paolino/dwarfstar-nix -c \
  ds4-server -m ./ds4flash.gguf --rocm --ssd-streaming --ctx 100000

Performance: --ssd-streaming and the GTT aperture

The 81 GB model exceeds the default ~62 GB GPU-visible (GTT) memory on Strix Halo, so the commands above use --ssd-streaming (experts streamed/cached, no kernel changes needed). It just works, at modest speed (~6 tok/s generation).

For full-residency speed, enlarge the GTT aperture via kernel params and reboot, then drop --ssd-streaming:

amd_iommu=off amdgpu.gttsize=126976 ttm.pages_limit=32505856 ttm.page_pool_size=32505856

(On NixOS: boot.kernelParams = [ "amdgpu.gttsize=126976" … ]; then nixos-rebuild switch + reboot.)

Other GPUs

Default target is gfx1151. The flake also exposes #ds4-gfx1100 (RDNA3 dGPU), or build for any arch:

nix build github:paolino/dwarfstar-nix#ds4 --override-input nixpkgs nixpkgs   # default gfx1151
# or override rocmArch in package.nix for another target

Flake outputs

• packages.x86_64-linux.default / .ds4 — the ROCm build (gfx1151)
• packages.x86_64-linux.ds4-gfx1100 — RDNA3 variant
• apps.x86_64-linux.default — runs ds4
• devShells.x86_64-linux.default — build env + rocminfo/rocm-smi

Binaries: ds4, ds4-server, ds4-bench, ds4-eval, ds4-agent, and ds4-download-model.

How the build works

ds4's strix-halo Makefile target drives everything through hipcc, whose bundled clang can't host-link on NixOS (no bare ld/crt/dynamic-linker). This flake splits the build: device code (*.cu) is compiled by hipcc (with the gcc toolchain, glibc headers and ROCm device-libs spelled out explicitly), and the final host link is done by the Nix-wrapped g++, which handles crt, the dynamic linker and rpaths. See package.nix.

Credits & license

ds4 / DwarfStar is by Salvatore Sanfilippo (antirez), BSD-2-Clause. This packaging (flake glue) is MIT — see LICENSE.