feat: add one-line install script for Linux/macOS

[OliverBryant]

Summary

Adds a one-line installer so Linux/macOS users can install Xinference with:

curl -fsSL https://raw.githubusercontent.com/xorbitsai/inference/main/scripts/install.sh | sh

The script installs Xinference as an isolated uv tool, so it does not touch the system Python and avoids PEP 668 (externally-managed-environment) errors. It bootstraps uv first if it is not already present.

Details

• New file scripts/install.sh — POSIX sh, supports Linux and macOS (Windows users are directed to pip in a virtualenv).
• Default installs only the base xinference package to stay small and cross-platform. Heavy extras such as vllm require Linux + CUDA and often fail on macOS/CPU-only machines, so they are opt-in.
• Environment variables:
  • XINFERENCE_EXTRAS — install optional backends, e.g. XINFERENCE_EXTRAS=all or XINFERENCE_EXTRAS=vllm,transformers.
  • XINFERENCE_VERSION — pin a version, e.g. XINFERENCE_VERSION=1.8.1 (a leading v is stripped).
• On success it prints how to start the server (xinference-local) and the Web UI URL (http://127.0.0.1:9997), and warns if the command is not yet on PATH.
• Documented in the README quick start and doc/source/getting_started/installation.rst.

Notes

• The docs reference the raw GitHub URL so the installer works immediately. If a short vanity domain (e.g. get.xinference.io) is set up later, the URL can be swapped.
• Validated sh -n syntax and exhaustively tested the package-spec assembly across XINFERENCE_EXTRAS / XINFERENCE_VERSION combinations. The actual uv tool install was not executed in CI-less local testing.

[gemini-code-assist]

Code Review

This pull request introduces a new one-line installation script (scripts/install.sh) for Linux and macOS that installs Xinference as an isolated tool using uv, and updates the README and documentation accordingly. The review feedback correctly identifies that the command uv tool update-shell referenced in the script does not exist, and provides a robust code suggestion to simplify the PATH warning logic.

[qinxuye]

Thanks for adding this. I think the scope needs to be clarified before this becomes the primary Quick Start installer. The base xinference package already depends on torch, so even the default installation is hardware-sensitive: on Linux we need to account for CPU, NVIDIA, AMD/ROCm, Intel GPU, and the driver-supported PyTorch backend. Please use hardware-aware backend selection such as uv --torch-backend auto, allow an explicit override, and keep hardware-specific backends such as vLLM on a clearly supported path. macOS also needs to distinguish Apple Silicon from Intel. Since uv tool creates an isolated environment, packages from an active Conda environment or virtualenv are not reused; the script should at least detect an existing xinference command and warn about PATH/environment conflicts. If the intended scope is only installing the CLI/server framework, the README should state clearly that common model backends are not installed and this should not replace the existing functional Quick Start. Otherwise, please provide explicit CPU/CUDA/ROCm/MLX modes, select a tested Python version, perform preflight checks, and add CI coverage that actually exercises the installer rather than only repository lint/tests.

[OliverBryant]

Thanks @qinxuye, this was the right call — the base package pulls in torch, so a naive uv tool install is indeed hardware-blind. I've reworked the installer to address each point:

Hardware-aware PyTorch. The script now creates a dedicated uv-managed virtualenv and installs via uv pip install --torch-backend "$XINFERENCE_BACKEND" .... --torch-backend is only available in the uv pip interface (not uv tool install), which is why I moved off the tool path. XINFERENCE_BACKEND defaults to auto (uv detects CUDA/ROCm/Intel driver and picks the matching index) and accepts explicit cpu / cuXXX / rocm / xpu overrides.

macOS Apple Silicon vs Intel. uname -m is checked; on x86_64 macOS the script warns that recent PyTorch no longer ships Intel-mac wheels. --torch-backend is skipped on macOS (single universal wheel).

Isolated env / conflict detection. Preflight now warns when an existing xinference is on PATH (it will be shadowed via ~/.local/bin), and when run inside an active virtualenv/conda env it warns that uv installs into a separate env and packages are not reused, pointing to pip install "xinference[all]" for in-place installs.

Scope is explicit. This installs only the base framework — no model backends. Both the README and the installation docs now state this clearly and position it as a convenience bootstrap, not a replacement for the existing pip install "xinference[all]" Quick Start, which is kept as the functional path. Extras remain opt-in via XINFERENCE_EXTRAS.

Pinned Python + preflight + CI. The venv pins Python (default 3.12, override via XINFERENCE_PYTHON). A new install-script.yaml workflow shellchecks the script and runs it end-to-end in CPU mode on Linux and macOS, asserting xinference-local --help and xinference --version work — so the installer itself is exercised, not just repo lint/tests.

I verified the flow end-to-end locally on macOS arm64 (installs base xinference, links the commands, xinference-local --help returns 0). vLLM/ROCm/XPU paths are documented but not exercised in CI (no such hardware on hosted runners). Let me know if you'd prefer a different default backend or Python version.