Software 45247 Published by

AMD shipped ROCm 10.2.0 on November 18, 2026, continuing its six-week release cycle with a stability-focused refresh instead of a major new feature. The upgrade bumps serving engines to vLLM 0.32.0 and SGLang 0.5.21 on Python 3.14, refreshes training frameworks like PyTorch 2.15, and carries AMD's claim that the Instinct MI355X trained DLRM v4 38.5% faster than NVIDIA's B300. It also broadens support to 18 new Ryzen processors and the Radeon PRO W7900D, tunes RCCL for eight-GPU Radeon servers, and trims library footprints with compressed catalogs while adding complex-math and storage I/O improvements. Ultimately, AMD is betting that predictable, low-friction updates build the trust CUDA has earned over a decade, though some newer tools like Hyperloom's Radeon support remain early access.





AMD ships ROCm 10.2.0, betting stability beats flash

AMD's ROCm 10.2.0 shipped on November 18, 2026, and it isn't trying to sell you on a headline feature. This is a refresh. The six-week release cycle AMD locked in under ROCm 10 held firm: 10.1 arrived October 5, and 10.2.0 followed the same schedule. AMD says the release "prioritizes stability," refreshing the frameworks, libraries, and tools developers already lean on so they can upgrade "with confidence."

ROCm is AMD's open answer to CUDA, and the version bump alone tells a story. AMD jumped its numbering from 7.14 straight to 10.0 in August 2026. That wasn't a typo or marketing gymnastics. The leap coincided with the full switch to TheRock, a new build and release system, plus the debut of ROCm.AI and its ROCm CLI and AMD Skills tools. That 10.0 release was pitched as a "ten-year milestone."

Shipping 10.2.0 on schedule, then, is itself the point.

On the serving side, this release moves vLLM up to 0.32.0 and SGLang to 0.5.21, both now running on Python 3.14. The training stack gets PyTorch 2.15 wheels alongside JAX 0.11.2, MIGraphX 2.19, and ONNX Runtime 1.29. One change that will bite people who automate installs: TheRock-built PyTorch wheels now pull ROCm in as a separate dependency instead of bundling it inside the package. Keep in mind that if your install script never expected that split, it may need a tweak.

There is one claim that's bigger than the rest. AMD points to its MLPerf 6.1 submission and says its Instinct MI355X trained the DLRM v4 benchmark in 38.5% less time than NVIDIA's B300, holding its own against the GB300 at 8-node scale. Those are AMD's own numbers, and MLPerf runs where the vendor drives the hardware always come with a grain of salt. Still, 38.5% on a workload people actually use is worth a look.

Rocm

Shrinking the friction

For what it's worth, the most interesting moves here are the ones nobody will write headlines about.

AMD's InferaSim lets you project how a serving config will perform from a handful of measured GPU runs, does it fit in memory, how many replicas does your latency target need, how should you split prefill from decode. Hyperloom gains power-budget utilities so it can optimize for watts per token rather than just tokens per second. AI Daily Post called this "a small feature with a big tell": the market is quietly shifting toward efficiency as the real deployment metric, and fleet operators feel that in the electricity bill.

The ROCm CLI v0.3.0 makes rocm install sdk pull the latest stable release by default, and rocm serve now adds an AIM container engine for "AMD Enterprise AI" alongside vLLM and Lemonade, plus a new /v1/messages endpoint with tool use and streaming. Newer still are AMD Skills, which let AI coding agents work directly with AMD's stack. The collection now covers 12 projects, and this build adds three skills for kernel work: writing FlyDSL kernels, analyzing their traces, and porting a CUDA project over via moat-cuda-to-rocm. You can pull them all with a single npx skills add amd/skills command.

On the math libraries, hipBLASLt gains single- and double-precision complex matrix multiplication (CGEMM and ZGEMM), matching across the MI300 and next-gen MI350 series. In a move that actually saves disk space, hipBLASLt and hipSPARSELt now ship compressed solution catalogs that shrink the footprint from roughly 3.8 GB down to about 194 MB. The libraries decompress on first use, so you pay that hit once. rocFFT adds JIT callbacks for multi-GPU transforms and will auto-switch to 64-bit indexing when problems exceed 32-bit addressing, which quietly prevents offset-overflow bugs on huge transforms.

Getting the basics working

Support breadth is where this release quietly does the most homework. AMD adds 18 new Ryzen processors built on the gfx1103 architecture already found in the Radeon 700M-series chips, spanning Ryzen 3, 5, 7, and 9 PRO and non-PRO SKUs, plus the Radeon PRO W7900D. Compatibility spans Windows and Linux with no custom builds required. It also adds RHEL 10.3, RHEL 9.9, and SLES 16.1.

RCCL gets tuned for single-node eight-GPU Radeon PCIe servers, validated with four-way tensor-parallel vLLM serving. That's a signal AMD wants data-center-style collective comms working on its mainstream Radeon line, not just Instinct. hipFile and AMD Infinity Storage gain PyTorch's load_storage and save_storage APIs, plus support in fio v3.43 and ior, and now back LVM and mdraid so you can abstract multiple devices as one.

A new unified rocsys command consolidates the Systems Profiler under one interface, and the Streaming Performance Monitor adds a device counting mode that samples even between kernel dispatches, where work previously went unmeasured. On the Physical AI front, the AMD Simulation Toolkit picks up four libraries, Quadrants, Genesis, Warp, and MuJoCo-Warp, positioning AMD beyond generative models into physics simulation.

The trust bet

Read against the ambition of 10.0 and the WSL2 momentum of 10.1, 10.2.0 is a consolidation release. It refreshes serving engines, trims library footprints by gigabytes, adds complex-math and sparse primitives, and smooths profiling and storage I/O. That's deliberate.

The strongest card AMD holds against CUDA isn't a single benchmark. It's predictability. Shipping vLLM, SGLang, and PyTorch wheels on a reliable six-week release cycle, with changes that mostly reduce friction rather than force rewrites, is how an ecosystem earns the trust engineers need to standardize on it. AMD is arguably competing on reliability more than raw performance at this point, and for a stack trying to break a decade of CUDA lock-in, that's arguably the smarter bet.

For what it's worth, treat a couple of the newer features with caution. Hyperloom's single-node Radeon support and hipFile's ior integration are early-access or freshly merged, so real-world maturity may lag the announcement. AMD also notes its standard disclaimer that information "may contain technical inaccuracies" and is subject to change. Validate against your own workloads before re-architecting anything.

The release is available now through AMD's normal channels. The full write-up, including companion pieces on the MoE training work in Primus and the individual Simulation Toolkit libraries, is on AMD's ROCm blog, where it's authored by Amy Wiebe, Adam Chan, Onil Gunawardana, John Tyler, and Saad Rahim. Head here to read it.