# Launch TRT-LLM server with 4B FP8 engine on port 5000 (no Medusa)
# TRTLLM_ENABLE_XQA_JIT=0 is mandatory — see hard_won_facts.md

$env:TRTLLM_ENABLE_XQA_JIT  = "0"
$env:TLLM_DISABLE_MPI        = "1"
$env:PYTORCH_ALLOC_CONF      = "expandable_segments:True"
$env:TENSORRT_PATH           = "D:\AI\apps\TensorRT-sdk-local"
$env:TRT_LIBPATH             = "D:\AI\apps\TensorRT-sdk-local\lib"
$env:CUDA_PATH               = "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.2"
$env:PATH                    = "D:\AI\apps\TensorRT-LLM\tensorrt_llm\libs;D:\AI\apps\TensorRT-sdk-local\lib;D:\AI\apps\TensorRT-sdk-local\bin;D:\AI\apps\TensorRT-LLM\.venv-3.11\Lib\site-packages\tensorrt_libs;D:\AI\apps\TensorRT-LLM\.venv-3.11\Lib\site-packages\torch\lib;C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.2\bin;C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.2\bin\x64;" + $env:PATH
$env:PYTHONPATH              = "D:\AI\apps\TensorRT-LLM"

$python = "D:\AI\apps\TensorRT-LLM\.venv-3.11\Scripts\python.exe"

Write-Host "=== Starting Josie 4B FP8 server on :5000 ===" -ForegroundColor Cyan
Write-Host "Engine: D:\AI\models\Josiefied-Qwen3-4B-fp8-engine"
Write-Host "Kill llama.cpp first if running (both need full VRAM)"
Write-Host ""

& $python "D:\AI\models\start_server_5000_fp8.py"