# ---------------------------------------------------------------------------
# CoderAI — vLLM high-concurrency backend, ISOLATED venv.
#
# vLLM pins its own torch/CUDA (e.g. torch==2.13.0 / cu13) which conflicts with the main
# coderai venv, so it runs in a dedicated virtualenv + managed subprocess (same isolation
# pattern as the OCR Paddle/Surya engines and ktransformers/SGLang). coderai launches
# `python -m vllm.entrypoints.openai.api_server` from this venv and proxies to it
# (codai/backends/vllm.py). Also serves Surya2 (a VLM) for the OCR subsystem.
#
#   python3 -m venv <dir> && <dir>/bin/pip install -r requirements-vllm.txt
#   (or set vllm.auto_build = true to have coderai build it on first use)
#
# vLLM's wheel pulls its exact torch/torchvision/torchaudio + CUDA runtime, so do NOT pin
# torch here — let vLLM choose the matching versions. Pin only vLLM itself.
# ---------------------------------------------------------------------------
vllm==0.27.1
