#!/usr/bin/env bash
set -euo pipefail
export HF_HOME=${HF_HOME:-/workspace/holo3/cache}
export VLLM_USE_V1=${VLLM_USE_V1:-1}
export VLLM_USE_DEEP_GEMM=${VLLM_USE_DEEP_GEMM:-0}
export VLLM_MOE_USE_DEEP_GEMM=${VLLM_MOE_USE_DEEP_GEMM:-0}
export VLLM_ENABLE_RESPONSES_API_STORE=${VLLM_ENABLE_RESPONSES_API_STORE:-1}
exec vllm serve /workspace/holo3/models/Holo-3.1-35B-A3B-NVFP4-bf16-head \
  --host 0.0.0.0 \
  --port 8000 \
  --served-model-name holo3-1-35b-a3b-mixed-nvfp4 \
  --trust-remote-code \
  --dtype auto \
  --gpu-memory-utilization 0.90 \
  --max-model-len 65536 \
  --enable-prefix-caching \
  --safetensors-load-strategy prefetch \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_xml