Spaces:
Paused
Paused
Commit History
-
feat(l1-coord+llm-fleet): LangGraph coordinator + 2026 SOTA CPU LLM refresh + Logfire OTel
Ship steps (all 4 completed per "DO all absolutely"): 1. Self-host LLM Space refresh (URLs preserved, weights bumped to 2026 SOTA): - qwen/: Qwen2.5-0.5B -> Qwen3-0.6B (36T training, /think mode) - llama/: Llama-3.2-1B -> Dolphin3.0-Llama3.2-3B (JSON/function-calling tuned) - gemma/: Gemma-2-2B -> Gemma-4-E2B (Apr 2 2026, native JSON) - hf-cpu-gemma4/: fallbacks now Phi-4-mini-instruct (MIT, 128K ctx) 2. Gateway wiring (scripts/arena/hf-llm-gateway/app.py): - Model IDs renamed to selfhost:{phi-4-mini,qwen3-0.6b,dolphin3-llama-3.2-3b,gemma-4-e2b} - FALLBACK_CHAINS rebuilt around new selfhost keys - Logfire OTel wrapping every call_llm dispatch; soft-dep if LOGFIRE_TOKEN unset - requirements.txt: +logfire>=2.0 3. L1 Coordinator on VM (scripts/coord/coord.py, NEW): - Magentic-One thin-orchestrator pattern (arXiv 2411.04468) - LangGraph StateGraph: scan -> decide -> dispatch -> record - 21 islands scanned in parallel (ThreadPoolExecutor, 8 workers) - …06e8cb6
-
feat(llm-fleet): deploy 3 quantized CPU LLM Spaces (Qwen2.5-0.5B + Llama-3.2-1B + Gemma-2-2B) + gateway wiring
- Nomos42/qwen25-05b-cpu — Qwen2.5-0.5B-Instruct GGUF Q4_K_M (~400 MB) - Nomos42/llama32-1b-cpu — Llama-3.2-1B-Instruct GGUF Q4_K_M (~800 MB) - Nomos42/gemma2-2b-cpu — gemma-2-2b-it GGUF Q4_K_M (~1.6 GB) Each Space: pure FastAPI (no Gradio), Docker SDK, llama-cpp-python, port 7860, OpenAI-compatible /chat/completions + /v1/chat/completions, root / health returns {model, ready}. Gateway (LBJLincoln26/llm-gateway) wired with 3 new selfhost: MODEL entries and appended as last-resort on every FALLBACK_CHAINS row.bfaf3ea