Update via HuggingBot

runtime-baseline.md

@@ -0,0 +1,45 @@
+# OpenJev CPU runtime — baseline Space spec
+
+The recommended zero-GPU Hugging Face Space configuration for
+**broadfield/openjev-cpu-run** (the one-click Duplicate target).
+
+| setting | value | why |
+|---|---|---|
+| SDK | **Docker** | we need llama.cpp binaries + a multi-process entrypoint (bash) |
+| Hardware | **CPU basic (free)** | zero GPU, by design (note: Docker Spaces on cpu-basic require a PRO subscription) |
+| Persistent storage | **30 GB** (≈ $0.50/mo) | `/data` holds the 16.2 GB GGUF + tokenizer |
+| Auto-sleep | any (e.g. 15 min) | wake-up reloads model from `/data` → ~1–2 min, not 20+ |
+| Visibility | public (default) | the API is meant to be called; flip private + `OPENJEV_TOKEN` to lock it |
+| Secrets | none required | optional `OPENJEV_TOKEN` via Space settings → will appear as an env var |
+
+## Approximate resource/memory math (Q4_K_M)
+
+| item | size |
+|---|---|
+| `OpenJev-Q4_K_M.gguf` on disk | 16.2 GB |
+| RAM to hold the model weights | ~16.2 GB |
+| KV cache (llama.cpp fork, 8-bit, 8192 ctx, 1 stream) | ~1–2 GB |
+| openjev-server + transformers tokenizer | ~1–2 GB |
+| **peak RSS** | **~18–20 GB** (fits the 16 vCPU / 32 GB free tier) |
+
+If the free tier's memory turns out tight (it runs 32 GB RAM / 16 vCPU per
+current info), drop to `LLAMA_CTX=4096` or use quant `OpenJev-Q5_K_M.gguf`.
+
+## Build-time budget
+
+The image does, in order:
+1. `apt` build tools — ~1 min
+2. clone + checkout llama.cpp `b4630` — ~1 min
+3. **CMake CPU build of llama-server** — **~15–35 min** on the build machine
+4. `pip install openjev-server` (transformers etc.) — ~2–4 min
+5. final image copy — ~1 min
+
+First **boot** then downloads the GGUF (~16.2 GB) into persistent `/data` —
+that's the long pole (20–50 min depending on HF Hub egress). Persisted
+afterward, so subsequent boots are 1–2 min.
+
+## Why entrypoint waits up to 15 min for llama-server
+
+A cold boot loads 16.2 GB off disk into RAM and runs llama.cpp's warmup; on a
+shared CPU that's minutes. The entrypoint polls `/v1/models` until ready, then
+starts openjev-server, so HF's health check never sees a crash while loading.