Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

40 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Accretion — the disk around DwarfStar. A consumer streaming platform that feeds ds4 experts from NVMe on demand, so a 300B-class model serves real work from hardware you can actually buy.

What it is

DwarfStar (ds4) is the engine; Accretion is the platform around it. The split is deliberate:

  • Engine improvements — correctness fixes, measured performance wins, model compatibility — flow upstream to antirez/ds4.
  • Platform genericity — everything a multi-user serving deployment needs that the engine shouldn't carry — lives here.

Status

v0.1 reached (release v0.1.0): one install command, browser console, model picker (browse + switch the served model, admin-token gated), point your harness at it. Seeded from the nexus-cw/ds4 platform branch.

Reference box: running in production on a GB10 (robo-dog) serving DeepSeek V4 Flash MXFP4 at 5-6 tok/s from a 156GB GGUF with a 70GB expert cache. Note: that box predates the install layout — its unit carries the model path inline, so console model switching stays disabled there (safe default) until it migrates to the env-file layout.

Install

Consumers never build code — download a release, run the installer. Current target: gb10-arm64-cuda (GB10 / DGX Spark). Other targets (generic arm64 CUDA, x86_64 CUDA, Apple/Metal) are planned; see docs/RELEASING.md.

# 1. Download the latest release tarball (and .sha256) from
#    https://github.com/nexus-cw/accretion/releases
tar xzf accretion-<version>-gb10-arm64-cuda.tar.gz
cd accretion-<version>-gb10-arm64-cuda

# 2. Install (idempotent; never overwrites an existing config)
sudo ./install.sh

# 3. Configure: set DS4_MODEL (and any tuning) then start
sudoedit /opt/accretion/etc/ds4-server.env
sudo systemctl enable --now ds4-server

# 4. Point your harness at it (Anthropic-compatible API)
export ANTHROPIC_BASE_URL=http://<host>:8000
export ANTHROPIC_API_KEY=none   # accepted but unused
claude   # or any Anthropic-API client

sudo ./uninstall.sh removes it (config kept unless --purge). Coming next: a web console (#27) and a model downloader with setup-and-optimize UX, so "pick a model" is a browser action rather than an env file.

Architecture sketch

  • engine/ — ds4 with expert streaming + CUDA expert LRU (subtree of nexus-cw/ds4 platform).
  • Platform layer roadmap:
    • Anthropic + OpenAI API surface
    • Honest capability endpoint (what the box can actually do, right now)
    • KV pinning + slot-affinity for large system prompts
    • Prefill-quantum interleaving for multi-user serving
    • Tiered I/O backend: fadvise / O_DIRECT / GDS, auto-selected
    • Model setup-and-optimize downloader
    • Go routing proxy

Relationship to upstream

Engine fixes go to antirez/ds4 first — see PRs 647, 659, 662, 664 — and we track upstream main continuously. See docs/UPSTREAM_LANES.md for the sorting rule and scripts/sync-engine.sh for the sync mechanics.

Attribution

The engine (engine/) is DwarfStar by Salvatore Sanfilippo (antirez), MIT licensed. The platform layer is copyright the Accretion contributors, also MIT. See LICENSE.

About

The accretion disk around DwarfStar — a consumer streaming platform that feeds ds4 experts from NVMe on demand, so 300B-class models serve real work from hardware you can actually buy.

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages