Phase 17 — Cheatsheet: Hardware Backends & Plugins

• Platform abstraction = one place for device/dtype/default-backend/capabilities.
• Engine asks the Platform; it never hardcodes CUDA.
• New hardware = out-of-tree plugin via entry points.
• CPU backend runs on a laptop (no paging/graph kernels), great for learning.

Key upstream files

• vllm/platforms/interface.py
• vllm/platforms/cuda.py
• vllm/platforms/cpu.py
• vllm/platforms/__init__.py
• vllm/plugins/

Full reference: 00-guide.md · 01-deep-dive.md