vLLM explains how it migrated to the LibTorch Stable ABI
AIvLLM describes how it moved its CUDA build to the LibTorch Stable ABI, so a new torch wheel can be imported without rebuilding vLLM. The post lists the migration steps, from assessing API gaps and moving meta functions to Python, to replacing unstable APIs, verifying changes, and checking binaries with torch-abi-audit. It also notes tradeoffs, including a small host-side overhead in some paths, and says that the vLLM CUDA wheel became ABI stable after its seven LibTorch-dependent build dependencies were migrated.
AIWhy it matters.
The post walks through the migration steps, tradeoffs, and verification tooling, showing how a project can keep its wheel usable across PyTorch upgrades without rebuilding.
Source details
vLLM Blog · vLLM Team, Red Hat PyTorch Team, and Meta PyTorch Team · vllm.aiPublished · added here
Original title: Migrating vLLM to the LibTorch Stable ABI