Apppublic

LisaMegaWatts/MonarchSLM

sourceHugging Facemitupdated 7mo agoView on Hugging Face

0likes

App README

MonarchSLM

A Monarch Mixer decoder-only model (sub-quadratic sequence mixing, RMSNorm, SwiGLU) trained on classical philosophy texts, implemented in Julia with Lux.jl. Serves an OpenAI-compatible API with streaming support.

Endpoints

—GET / — Health check and model info
—GET /v1/models — List available models
—POST /v1/chat/completions — Generate text (supports streaming, top-k, top-p)

Usage

bash

# Non-streaming
curl -X POST https://your-space.hf.space/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"messages": [{"role": "user", "content": "the nature of"}], "max_tokens": 200}'

# Streaming
curl -X POST https://your-space.hf.space/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"messages": [{"role": "user", "content": "the nature of"}], "stream": true, "temperature": 0.7, "top_k": 40}'

Architecture

—Model: ~5M params, 256d embed, 8 layers, 8 Monarch heads
—Sequence mixing: Multi-head Monarch Matrix (sub-quadratic) + Causal Depthwise Conv + Learned Gate
—Tokenizer: BPE (2000 tokens)
—Framework: Lux.jl (explicit parameter/state management)
—Normalization: RMSNorm (pre-norm)
—Feed-forward: SwiGLU activation
—Weight tying: Shared embedding/output projection
—Inference: CPU-only, no Lux dependency at runtime (pure NNlib)

Environment Variables

—HF_REPO — HuggingFace model repo (default: LisaMegaWatts/MonarchSLM)
—PORT — Server port (default: 7860)