Oossa

vLLM 0.34.0 adds EFA support to its CUDA images

The release adds AWS EFA userspace to the default CUDA image and configures the -efa image to route NCCL over EFA by default.

NoteBy Published by Oossa: 1 min read

The vLLM project released version 0.34.0 on November 17, 2026, with contributions from 412 people. Its default CUDA image now includes AWS Elastic Fabric Adapter (EFA) userspace libraries, while the separate -efa image adds DeepEP built for EFA and routes NCCL traffic over EFA by default.

The release also provides wheels for CUDA 13.0 on x86_64 and arm64, plus CPU and XPU builds.

Why it matters

For teams running vLLM on AWS, the image changes make EFA components available by default in the CUDA image, but the release notes do not quantify any performance gains.

Was this article useful?
Share

Read next

Oossa · Newsletter

The week in AI, explained

Every Monday: the stories worth knowing, in plain language. Free, no spam.