AMD announced ROCm 10.2.0 on November 18, 2026, describing it as a stability-focused release. Its developer-facing changes include updated serving frameworks, storage and profiling additions, RCCL tuning for a specific eight-GPU setup, and support for 18 additional gfx1103 Ryzen processors. AMD’s release announcement details the scope of those changes.
Match each release change to its stated workload: the RCCL tuning targets a particular Radeon server configuration, while the Ryzen support covers gfx1103 processors and named frameworks.
Which framework versions and tools does ROCm 10.2 include?
AMD lists vLLM 0.32.0 and SGLang 0.5.21, with both serving frameworks running on Python 3.14. The distribution options differ: AMD provides vLLM in prebuilt Docker images and Python wheels, while SGLang is available in Docker images.
The official vLLM Docker images package AITER, so a separate manual installation is not needed for those images. AMD lists prebuilt AITER kernels for Instinct MI300X, MI325X, MI350X and MI355X. On Radeon and Ryzen AI, the kernels compile automatically on first use.
What workloads does the RCCL tuning target?
AMD describes additional RCCL tuning for single-node Radeon PCIe servers with eight GPUs, aimed at tensor-parallel serving. The company says it validated the tuning with four-way tensor-parallel vLLM serving. That scope is relevant when assessing the change for a matching deployment; it describes a specific server and workload.
What changed in storage and GPU libraries?
AMD says hipFile and AMD Infinity Storage add support for LVM and Linux mdraid. The release also includes integration work involving PyTorch, fio and ior. For PyTorch storage workflows, AMD names the load_storage and save_storage APIs for tasks such as loading model weights and saving checkpoints.
For GPU libraries, AMD reports that hipBLASLt adds complex GEMM support for MI300 and MI350 series architectures, identified as gfx942 and gfx950. The release also adds strided batched triangular solvers to rocSPARSE. In rocFFT, transforms whose problem sizes exceed 32-bit addressing limits use 64-bit indexing.
How does profiling and Ryzen support change?
ROCm 10.2 introduces rocsys, an experimental unified command-line entry point for the ROCm Systems Profiler. AMD also says the profiler can attach from a host to a running container without manual synchronization of runtime libraries.
AMD says the release validates support for 18 additional Ryzen processors based on gfx1103 across Windows and Linux for HIP, PyTorch and vLLM. The support claim applies to those named frameworks and operating systems; the release announcement describes it as validated compatibility.