entrypoint.sh

#!/usr/bin/env bash
# ─────────────────────────────────────────────────────────────────────────────
# OpenJev zero-GPU entrypoint (Docker Space).
#   1. Download the chosen GGUF once into /data (HF persistent storage) — the
#      16.2 GB load survives restarts/sleeps, so wake-ups are fast.
#   2. Fetch the OpenJev tokenizer (needs tokenizer.json + chat template).
#   3. Start llama.cpp `llama-server` on CPU with logprobs enabled.
#   4. Start the OFFICIAL openjev-server, its `vllm` backend pointed at the
#      llama.cpp OpenAI endpoint (top-logprobs fallback path, no code changes).
#      openjev-server binds $OPENJEV_PORT (7860) - the HF Spaces ingress.
# ─────────────────────────────────────────────────────────────────────────────
set -euo pipefail

log() { printf '\n\033[1;34m[entrypoint]\033[0m %s\n' "$*"; }

# config
MODEL_REPO="${MODEL_REPO:-openjev/openjev-GGUF}"
MODEL_FILE="${MODEL_FILE:-OpenJev-Q4_K_M.gguf}"
TOKENIZER_REPO="${TOKENIZER_REPO:-openjev/openjev}"
LLAMA_PORT="${LLAMA_PORT:-8080}"
OPENJEV_PORT="${OPENJEV_PORT:-7860}"
LLAMA_CTX="${LLAMA_CTX:-8192}"
LLAMA_MAX_LOGPROBS="${LLAMA_MAX_LOGPROBS:-64}"
LLAMA_PARALLEL="${LLAMA_PARALLEL:-1}"
LLAMA_THREADS="${LLAMA_THREADS:-}"
LLAMA_THREADS_BATCH="${LLAMA_THREADS_BATCH:-}"
MODEL_DIR="${MODEL_DIR:-/data/models}"
MODEL_PATH="${MODEL_DIR}/${MODEL_FILE}"
TOKENIZER_DIR="${TOKENIZER_DIR:-/data/tokenizer}"

mkdir -p "$MODEL_DIR" "$TOKENIZER_DIR"

# ── 1. model (one-time download; persists in /data across restarts) ─────────
if [ ! -s "$MODEL_PATH" ]; then
  log "downloading $MODEL_REPO/$MODEL_FILE -> $MODEL_DIR  (one-time, ~16 GB, persistent)"
  HF_HUB_ENABLE_HF_TRANSFER="${HF_HUB_ENABLE_HF_TRANSFER:-1}" python - <<'PY'
import os
from huggingface_hub import hf_hub_download
repo, fname, d = os.environ["MODEL_REPO"], os.environ["MODEL_FILE"], os.environ["MODEL_DIR"]
print("hf_hub_download:", repo, fname)
p = hf_hub_download(repo_id=repo, filename=fname, local_dir=d, resume_download=True)
print("downloaded to", p)
PY
else
  log "model already on disk: $MODEL_PATH ($(du -h "$MODEL_PATH" | cut -f1))"
fi

# ── 2. tokenizer snapshot (openjev-server needs tokenizer.json + chat template)
if [ ! -e "$TOKENIZER_DIR/tokenizer.json" ]; then
  log "fetching tokenizer from $TOKENIZER_REPO"
  python - <<'PY'
import os
from huggingface_hub import snapshot_download
print(snapshot_download(repo_id=os.environ["TOKENIZER_REPO"], local_dir=os.environ["TOKENIZER_DIR"],
                        allow_patterns=["tokenizer*", "vocab*", "merges*", "*.json", "config*", "special_tokens_map*", "*.jinja"]))
PY
else
  log "tokenizer already present in $TOKENIZER_DIR"
fi

# ── 3. llama.cpp server (CPU) ───────────────────────────────────────────────
LLAMA_ARGS=(
  --host 127.0.0.1 --port "$LLAMA_PORT"
  --model "$MODEL_PATH"
  --ctx-size "$LLAMA_CTX"
  --n-predict 1
  --parallel "$LLAMA_PARALLEL"
  --logprobs "$LLAMA_MAX_LOGPROBS"
  --alias openjev
  --log-file /tmp/llama-server.log
  --no-warmup
)
[ -n "$LLAMA_THREADS" ]       && LLAMA_ARGS+=(--threads "$LLAMA_THREADS")
[ -n "$LLAMA_THREADS_BATCH" ] && LLAMA_ARGS+=(--threads-batch "$LLAMA_THREADS_BATCH")

log "starting llama-server (CPU) on 127.0.0.1:$LLAMA_PORT"
llama-server "${LLAMA_ARGS[@]}" &
LLAMA_PID=$!

# wait until llama.cpp accepts requests (up to 15 min on a cold boot)
python - <<'PY'
import os, sys, time, urllib.request
url = f"http://127.0.0.1:{os.environ['LLAMA_PORT']}/v1/models"
for _ in range(450):
    try:
        with urllib.request.urlopen(url, timeout=10) as r:
            if r.status == 200:
                print("llama-server is ready"); sys.exit(0)
    except Exception:
        pass
    time.sleep(2)
print("llama-server did not become ready in time"); sys.exit(1)
PY

# ── 4. official openjev-server (vllm backend -> llama.cpp fallback path) ────
export OPENJEV_BACKEND=vllm \
       OPENJEV_VLLM_URL="http://127.0.0.1:${LLAMA_PORT}/v1" \
       OPENJEV_MODEL="$TOKENIZER_DIR" \
       OPENJEV_SERVED_MODEL_NAME="${OPENJEV_SERVED_MODEL_NAME:-openjev}" \
       OPENJEV_HOST=0.0.0.0 \
       OPENJEV_PORT="$OPENJEV_PORT" \
       OPENJEV_PROFILE="${OPENJEV_PROFILE:-openjev}"

cleanup() { kill "$LLAMA_PID" 2>/dev/null || true; }
trap cleanup EXIT INT TERM

log "openjev-server listening on 0.0.0.0:$OPENJEV_PORT  (UI + /v1/systemone + /docs + /healthz)"
exec openjev serve --force