#!/bin/bash
# vLLM serve script for ThinkingCap-Qwen3.6-27B-FP8 on dankclaw (ROCm)
# Run via systemd service: vllm-thinkingcap.service

set -euo pipefail

# Environment
export HSA_OVERRIDE_GFX_VERSION=11.5.1
export HF_HOME="/home/dank/.cache/huggingface"

# vLLM virtual environment — use direct paths, no activate script
VENV="/home/dank/vllm-env"
export PATH="${VENV}/bin:${PATH}"

# Model
MODEL="bottlecapai/ThinkingCap-Qwen3.6-27B-FP8"

# Serve
exec "${VENV}/bin/vllm" serve "${MODEL}" \
    --port 8000 \
    --host 0.0.0.0 \
    --kv-cache-dtype fp8 \
    --gpu-memory-utilization 0.80 \
    --max-model-len 32768 \
    --max-num-batched-tokens 2048 \
    --max-num-seqs 32 \
    --reasoning-parser qwen3 \
    --language-model-only \
    --dtype auto