AMD ROCm 7.14.0 Drops With Modular "TheRock" Architecture and Major Framework Updates
The latest production release shifts the build system, adds seven new Ryzen AI APUs, and pushes PyTorch to 2.12.
AMD has pushed ROCm 7.14.0 into the wild, and this isn't just another routine version bump. The release marks a hard architectural pivot to "TheRock," a completely modular build and release system that lets you install only the runtime and developer components you actually need.
Keep in mind that the ongoing fragmentation of GPU compute software has made modular installs a long-overdue fix. The new approach trims the footprint by dropping legacy dependencies and swapping them for use-case-specific SDKs. You get AI, data science, or HPC toolchains on demand.
Framework Updates and HIP Parity
PyTorch jumps to 2.12.0, JAX hits 0.10.0, and vLLM lands at 0.23.0. If you're running inference workloads, AMD explicitly notes that the vLLM packages are now deployment-ready. The HIP layer isn't sitting idle either. New batch memory management APIs cut down on driver overhead, and execution context support finally gives you proper compute resource partitioning on a single card.
The API changes are practical. You can now discard or prefetch multiple memory ranges in a single call instead of spinning up dozens of driver requests. Faster HIP graph replay eliminates the map and unmap gaps that previously stalled kernel chains. It's a meaningful step toward closing the CUDA parity gap, though the transition won't happen overnight for legacy codebases.
On the profiling side, roctracer is officially out, replaced by rocprofiler-sdk as the native backend for PyTorch Profiler. AMD also slips in a beta flag for Streaming Performance Monitors, which finally gives you time-resolved hardware counters instead of single aggregated values per kernel. It only runs on MI300X through MI355X series hardware right now. The docs are blunt about stability risks. Don't spin this up in production.
Client-Side Reach and Cloud Partitioning
For the first time, ROCm officially supports a handful of client-side Ryzen AI APUs. That includes the Ryzen AI MAX+ PRO 495, MAX PRO 485 and 490, plus several 430, 435, and 445 variants across the gfx1151 and gfx1153 architectures. This is interesting because it pushes ROCm beyond server racks and into high-end workstations.
On the cloud side, AMD introduces multi-VF partitioning for the MI355X and MI350X. You can now split compute resources using DPX or CPX modes with NPS2 memory partitioning. If you're managing GPU clusters, that means tighter allocation without buying extra hardware.
Operating system support gets a fresh coat of paint. You'll find validated builds for Ubuntu 26.04, RHEL 10.2 and 9.8, SLES 16, and Debian 13. Driver versions 31.40.0 and 30.30.x cover the latest Instinct and Radeon PRO cards. Binaries are waiting at repo.radeon.com/rocm/apt/7.14.0 for Ubuntu and Debian, or the equivalent EL and SUSE paths if you're on enterprise Linux. Head to the GitHub release notes for the full compatibility matrix and the transition guide if you're migrating from the legacy stack.
The modular shift is a necessary evolution for an open compute platform that keeps growing. Whether TheRock actually smooths out the installation friction remains to be seen. ROCm isn't stopping here. Keep your eyes on the next preview releases as AMD works through the remaining SPM stability quirks.
Head here for the release notes and source code.
