AMD has rolled out its ROCm 7.14 software update, signaling a major shift in how the company packages and distributes its open source AI stack. According to the official AMD release documentation, the headline change is the production debut of TheRock. This is a rebuilt, automated build and release system designed to clean up the delivery pipeline and solve the packaging headache that has historically plagued the platform.
The transition is complete. The preview phase spanning versions 7.9 through 7.13 has ended, and ROCm 7.14 establishes the new baseline for future software releases. AMD is advising anyone still running version 7.2 or older to migrate immediately to avoid compatibility issues. The updated setup introduces a Core SDK containing the essential files most developers need, alongside separate packages for fields like computer vision and data science. The entire build process now runs on CMake, allowing the team to push stable nightly updates across both Windows and multiple Linux distributions.
Silicon compatibility has expanded across the board. The update brings production support to the AMD Instinct MI350 PCIe GPU family. Operating system coverage now matches the MI350X and MI355X, with new validation for SLES 16, Debian 13, and Red Hat Enterprise Linux 9.8. If you run container setups, the MI350P is now ready for Vanilla Kubernetes on Ubuntu and Red Hat OpenShift v4.21. On the consumer side, developers can now run the complete ROCm software stack on Ryzen AI MAX PRO processors, including the 495, 490, and 485 models.
Scaling across hardware is also getting easier. AMD has validated systems running 2, 4, or 8 Radeon AI PRO GPUs to help developers train larger models. On the software side, the Systems Profiler and Compute Profiler now run on Ryzen AI MAX PRO, while the Compute Profiler adds support for Strix Halo and Strix Point. Telemetry is also improved, with the AMD SMI command line tool showing GPU temperature, fan speed, power, and memory metrics on Ryzen AI setups.
Software frameworks have been updated to match current industry standards. The 7.14 update adds support for PyTorch 2.12 and JAX 0.10.0, while TensorFlow validation now covers versions 2.21, 2.20, and 2.19.1. If you run open source models, the company has added optimized configurations for Llama 3.1 8B Instruct, Whisper Large v3, and Qwen3.6 with vLLM on Radeon Linux. For creative tools, ComfyUI support now extends to Radeon RX 7000 and 9000 series cards.
The enterprise software layer has also received an update. SGLang, 1 of the most popular frameworks for serving large language models, is now supported on Radeon hardware for the first time. This gives businesses running multi user chatbots and agentic AI tools the flexibility to run their workloads on local workstations instead of relying solely on expensive data center chips.
The underlying math libraries received several technical updates. libhipcxx is now part of the Core SDK, and rocBLAS adds compilation support for the SPIR V open standard. For sparse matrix math, hipSPARSE adds Block Sparse Row format support to its vector and matrix multiplication tools to help close the feature gap with proprietary NVIDIA alternatives. Dense linear algebra libraries like hipBLASLt and hipTensor also received updates, with hipTensor now running on Windows and supporting RDNA 3 and RDNA 4 silicon.
Developers also get better control over hardware resources. This release introduces HIP Execution Contexts, which let you divide GPU compute power across different tasks at the same time instead of letting them fight over the whole chip. There is also hipFile, AMD direct storage library, which is now ready for production. It lets applications bypass the CPU completely, moving data directly between NVMe storage and GPU memory over fabrics or RDMA networks to drop latency.
Virtualization support has widened. The MI350 series now runs on VMware ESXi 9.1, and SR IOV configurations can support up to 64 virtual functions per node. For installation, a new Runfile installer lets you install the software without dealing with an OS package manager. This tool auto detects local GPUs and allows you to install multiple ROCm versions side by side.
