[Feature]: Improved Broadcom NIC support

Suggestion Description

Currently, AMD’s official ROCm container images provide limited support for bnxt modules. This has been a friction point for customers who are used to base images working out-of-the-box on the Nvidia side with Mellanox.

Current setup docs:
https://rocm.docs.amd.com/en/latest/how-to/rocm-for-ai/system-setup/multi-node-setup.html

This limits the usability of ROCm base containers for distributed workloads, as users must manually install and configure bnxt drivers or rebuild container images to support RoCE, which can require additional CI work to support different deployments.

We understand that supporting multiple RoCE implementations (Broadcom, etc.) may be non-trivial. However, even partial out-of-the-box support or the ability to select the target version via an ENV would significantly improve the usability and portability of ROCm containers for multi-node workloads.

Operating System

No response

GPU

No response

ROCm Component

No response

Start with the linked ROCm multi-node setup documentation and trace how the container images currently handle bnxt modules and RoCE support. Done means the issue has a concrete implementation scope that provides Broadcom support out of the box or allows the target version to be selected via an ENV.