GitHubHelp

Heretek-AI/chlorine-server

Clean-room AGPL-3.0 reimplementation of the halogen 0.1.3 inference engine for Qwen3.8-27B on AMD Strix Halo (gfx1151) — same wire protocol and OpenAI-compatible API. Not affiliated with Peonist.

★ 16Forks 4HIPGitHub ↗Compare
agpl-3-0amd-gpuclean-roomcpp23gfx1151halogenhipllm-inferenceopenai-apiqwen3reimplementationrocmstrix-halo

README

chlorine-server

An independent, AGPL-3.0 reimplementation of the halogen 0.1.3 inference engine for Qwen3.8-27B on AMD Strix Halo (gfx1151) — clean-room rebuilt from reverse-engineering notes in docs/halogen/.

  • Same wire protocol as halogen's engine (PING/INFO/CSTAT/GEN/X on the token port) and the same OpenAI-compatible serving API on :8731.
  • Loads .hgn checkpoints (see converter/ and docs/halogen/CHECKPOINT-FORMAT.md).
  • Greedy token streams match the original engine on identical weights.

Not affiliated with Peonist or halogen. The original engine is closed source and lives at peonist-ai/halogen-server; its EULA permits reverse engineering and interoperability. This codebase is written from behavioral specifications only.

Status

Work in progress — see docs/halogen/REBUILD-PLAN.md.

Component State
docs/halogen/ RE notes + specs in progress
engine/ C++23 host scaffolded
engine/kernels/ HIP gfx1151 scaffolded
api/ OpenAI front-end scaffolded
converter/ safetensors → .hgn scaffolded

Weights

Never in this repo. Bring your own: the halogen .hgn checkpoints are published separately (peonist-ai/halogen-qwen3.8-27b on Hugging Face), or convert your own Qwen3.8-27B safetensors with converter/.

License

AGPL-3.0 — see LICENSE. Model weights are governed by their own licenses and are explicitly excluded from this repository's terms.

Contributors

johnzzw

Issues