#!/usr/bin/env bash # sensenova-u1.5-8b-gb10 entrypoint. Runs inside the container; takes no arguments. # Port of gx10 ~/code/vlmtest/machines/gx10/start_sensenova_u1.sh: the same server # (vlmtest/sensenova_u1_server.py, copied unchanged) with the same env, on pinned weights mounted # read-only at /weights/model. Same server as gx10, the vendor package baked into the image (it self-registers its model with transformers, so no trust_remote_code). set -eu export HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True export MODEL_DIR="/weights/model" export SERVED_MODEL_ID="sensenova-u1.5-8b-mot" export SENSENOVA_LORA="/weights/loras/SenseNova-U1.5-8B-MoT-LoRA-8step.safetensors" export PORT="$M64_PORT" echo '{"phase":"loading","detail":"loading SenseNova-U1.5-8B-MoT (8-step)"}' >> "$M64_STATUS" # The server loads its model before it listens (or answers /health with "loading"): # report ready once /health says "ok". python3 - <<'PY' & import json, os, time, urllib.request t0 = time.time() url = "http://127.0.0.1:%s/health" % os.environ["M64_PORT"] opener = urllib.request.build_opener(urllib.request.ProxyHandler({})) while True: try: r = opener.open(url, timeout=3) if r.status == 200 and json.loads(r.read() or b"{}").get("status", "ok") == "ok": break except Exception: pass time.sleep(5) with open(os.environ["M64_STATUS"], "a") as f: f.write(json.dumps({"phase": "ready", "detail": "loaded in %d s" % (time.time() - t0)}) + "\n") PY exec python3 -u /capsule/sensenova_u1_server.py