README.md

metadata

title: The Abyss
emoji: 🌀
colorFrom: purple
colorTo: green
sdk: docker
app_port: 7860
pinned: false

The Abyss — backend (API)

FastAPI inference server for Abyss-1, a ~50M-parameter language model trained from scratch (own byte-level BPE tokenizer; RoPE · RMSNorm · SwiGLU decoder), with streaming chat + a ReAct agent, JWT user accounts, a self-hosted safety/moderation layer, and sliding-window rate limiting.

This Space is assembled by deploy/hf-space/assemble.py in the project repo, which stages server/, ml/ (model + tokenizer) and agent/ next to this Dockerfile. The model weights are pulled from the HF Hub at boot (set ABYSS_HF_REPO) rather than baked into the image.

Configuration (Space → Settings)

Variables

• ABYSS_HF_REPO — HF model repo holding the checkpoint, e.g. your-name/abyss-1
• ABYSS_HF_CHECKPOINT — (optional) checkpoint filename (default agent_ckpt.pt)
• ABYSS_CORS_ORIGINS — (optional) comma-separated origin allowlist (default *)

Secrets

• SECRET_KEY — JWT signing key (use a long random value)
• ABYSS_ADMIN_KEY — unlocks GET /v1/admin/moderation

Leave ABYSS_API_KEYS unset to keep the API open to everyone.

Endpoints

GET /health · POST /auth/register · POST /auth/login · GET /auth/me · POST /v1/chat/completions (SSE) · POST /v1/agent/chat (SSE) · GET /v1/conversations · POST /v1/moderations

See DEPLOY.md in the repository for the full step-by-step runbook.