#!/usr/bin/env bash
# qwen38-flash-next-basalt-5090 entrypoint. Runs inside the container; takes no arguments.
# Basalt serves an OpenAI-compatible API on $M64_PORT. The pack is mounted read-only at
# /weights/basalt; nothing is fetched at runtime.
set -eu
echo '{"phase":"loading","detail":"loading the Basalt pack"}' >> "$M64_STATUS"

# Engine settings live in one config file (the pack carries no flags; server settings are the flags below): hal5090's
# start_qwen38_flash_next_basalt_384k.sh. Past the trained 262144 tokens the RoPE is
# stretched with YaRN by 393216/262144 = 1.5.
cat > /scratch/basalt_config.json <<'JSON'
{
  "engine": {
    "expert-cache": "auto",
    "prefill": "auto",
    "spec": 7,
    "spec-min-p": 0.5,
    "max-context": 393216,
    "kv": "int8",
    "resident-experts": true,
    "vram-reserve-mib": 700,
    "rope-scaling": "yarn",
    "rope-scale": 1.5,
    "yarn-orig-ctx": 262144
  }
}
JSON

# 0.0.0.0 inside the container's own network: only the m64 gate reaches it. The front calls
# it as Host "capsule", which Basalt's DNS-rebinding check must accept.
exec /opt/basalt/bin/basalt serve --config /scratch/basalt_config.json \
  --model /weights/basalt/Qwen3.8-FN-GSQ-RCO-IQ3_S.basalt \
  --alias qwen3.8-flash-next-iq3_s-basalt-384k \
  --vision dynamic \
  --host 0.0.0.0 --port "$M64_PORT" --allowed-host capsule