Static Linter tool EFA/NCCL container recipes

[gyaneshhere]

Purpose

Static Linter tool for EFA/NCCL container recipes

Changes

2.ami_and_containers/tools/
├── README.md ← README.md
└── efa-nccl-doctor/
├── README.md ← README.md
├── efa-nccl-doctor.sh
└── selftest.sh

[gyaneshhere]

ami_and_containers/tools/ currently holds two tools
ec2md (IMDSv2 metadata walker) and sagemaker_ftp/find-ftp.sh (Flexible Training Plan finder).
There's no README for tools/ itself. So I added a README.md for tools

The repo contains 36 Dockerfiles that build EFA-enabled distributed-training containers. Across them:

6 distinct EFA_INSTALLER_VERSION values (1.29.1, 1.35.0, 1.37.0, 1.47.0 ×6, 1.48.0 ×4, 1.49.0 ×2)
6 distinct AWS_OFI_NCCL_VERSION values with inconsistent naming — 1.12.1-aws, v1.7.3-aws, v1.13.2-aws, 1.18.0, v1.19.0 — because upstream dropped the -aws suffix at v1.14.0
Only 15 of 36 pin an EFA version at all, despite CONTRIBUTING.md explicitly requiring fixed versions over latest

Every one of these misconfigurations fails in the way that matters most, the image builds, the job launches, and NCCL silently falls back to TCP.

What I built

efa-nccl-doctor.sh — a static linter for EFA/NCCL container recipes, with 11 checks (EFA001-004, NCCL001-004, ENV001-002, PIN001), an --inventory mode for spotting drift, --json, --strict for CI, and an offline default.