StorageReview.com

AMD ROCm 10.2 Brings hipFile to PyTorch, LVM, and mdraid Volumes, and Adds 18 Ryzen Processors

AI  ◇  Enterprise

AMD has released ROCm 10.2.0, a stability-focused update to its open GPU software stack that refreshes the serving frameworks, libraries, and tools developers already run on Instinct, Radeon, and Ryzen hardware. For storage teams, hipFile and AMD Infinity Storage now plug into PyTorch, fio, and ior, and they work on volumes built with LVM and Linux mdraid. The release also brings AITER to Radeon and Ryzen AI, adds 18 Ryzen processors, and introduces an experimental unified profiler command.

AMD Instinct MI350P being lifted out of its PCIe riser in the Dell PowerEdge R7725, with the Instinct MI350P label on the passive cooling shroud facing the camera

Serving Builds, AITER, and Python 3.14

ROCm 10.2.0 moves AMD’s serving builds to vLLM 0.32.0 and SGLang 0.5.21, and both frameworks now run on Python 3.14. vLLM ships as prebuilt Docker images and Python wheels, while SGLang ships as Docker images. The official vLLM images also make the AI Tensor Engine for ROCm (AITER) easier to adopt, since AITER now runs on Radeon GPUs and Ryzen AI processors with no separate manual installation. Instinct MI300X, MI325X, MI350X, and MI355X GPUs get prebuilt AITER kernels, and on Radeon and Ryzen AI the kernels compile automatically the first time they are used.

MLPerf 6.1 and Primus MoE Training

AMD points to its latest MLPerf 6.1 training submission, which included five optimized recipes and a new entry for GPTOSS-20B. On the DLRM v4 benchmark, AMD says the Instinct MI355X trained in 38.5% less time than the NVIDIA B300 and was competitive with the NVIDIA GB300 at eight-node scale. These are vendor results, and we did not run or audit them.

Primus targets Mixture-of-Experts (MoE) training with optimized FP8 and MXFP8 kernels and a new fused MoE Megakernel. On DeepSeek-V3, AMD reports better-than-linear weak scaling from 256 to 1,024 Instinct MI355X GPUs with per-scale tuning.

Framework Updates

The release adds PyTorch 2.15 wheels, JAX 0.11.2, MIGraphX 2.19, and ONNX Runtime 1.29. PyTorch gains FlyDSL kernels for the MI350 and Tensor Data Mover (TDM) support for the MI450, and ROCm 10.0 wheels are now featured on the get-started page. PyTorch packages built by TheRock now install ROCm as a separate dependency and no longer bundle it inside the PyTorch package, a packaging change that teams pinning container images or build pipelines will need to account for.

AMD Instinct MI350P PCIe accelerator card shown without its cooler, with the GPU package at the center of the board

Deployment and Optimization Tools

ROCm InferaSim projects how a serving configuration will perform from a few short, measured GPU runs. It uses the CPU to estimate whether a configuration fits in GPU memory and how many replicas a latency target requires, and it helps decide how to split prefill (prompt processing) from decode (token generation).

Hyperloom, AMD’s agentic system for optimizing inference workloads on its GPUs, adds utilities for setting power budgets, so tuning can target efficiency as well as throughput. Early access support for single-node Radeon AI Pro R9700 systems extends Hyperloom beyond Instinct.

ROCm CLI v0.3.0 makes rocm install sdk install the latest stable ROCm, now 10.2.0, by default. rocm serve adds an AIM container engine for AMD Enterprise AI alongside vLLM and Lemonade, plus a new /v1/messages endpoint with tool use and streaming.

AMD Skills now covers 12 AMD projects, including Lemonade, Quark, and ZenDNN, so AI coding assistants can work with them directly. Three new skills cover kernel work and porting: flydsl-kernel-authoring, flydsl-kernel-trace-analysis, and moat-cuda-to-rocm, which lets Moat port a CUDA project to ROCm and HIP. Developers add them with npx skills add amd/skills.

RCCL Tuning and Ryzen Support

ROCm 10.2.0 adds RCCL tuning, including collectives, for single-node Radeon PCIe servers with eight GPUs. The work targets tensor-parallel serving, where RCCL handles collective communication among the GPUs sharing a model, and AMD says it validated the tuning with four-way tensor-parallel vLLM serving on eight-GPU Radeon servers.

ROCm also adds support for 18 AMD Ryzen processors built on the same gfx1103 architecture as the existing Radeon 700M Series. AMD says developers get Windows and Linux compatibility with no custom builds or special configurations, backed by validated gfx1103 support across HIP, PyTorch, and vLLM.

AMD Infinity Storage and hipFile

Building on the fast-path work in ROCm 10.1, AMD has extended AMD Infinity Storage (AIS) and hipFile to PyTorch, fio, and ior. PyTorch developers can call the load_storage and save_storage APIs, exposed through PyTorch’s native Python API, to move large tensors between GPU memory and storage. AMD says this brings hipFile’s bandwidth gains to model weight loading, checkpointing, and other I/O-heavy work.

hipFile support has landed in fio v3.43 and in the main branch of ior, which gives storage teams two familiar benchmarks for testing the storage-to-GPU path. hipFile also now works with volumes managed by Logical Volume Management (LVM) and with RAID sets built on the Linux kernel’s mdraid subsystem. AMD describes the change as a step toward compatibility with production storage infrastructure built around its GPUs.

Math Libraries

hipBLASLt adds host and client infrastructure for single- and double-precision complex matrix multiplication (CGEMM and ZGEMM). The initial release offers functional parity across the Instinct MI300 and MI350 Series (gfx942 and gfx950), and rocBLAS now routes complex GEMM calls to hipBLASLt. In Debian and RPM packages, solution-selection catalogs install as zlib-compressed .dat.zlib files, which cuts their disk footprint from about 3.8GB to about 194MB. Catalogs decompress on first use, which can add initial overhead, and AMD says later calls are unaffected.

rocSPARSE introduces strided batched triangular solvers (rocsparse_sptrsm) that dispatch thousands of independent matrix solves in a single GPU call. A new memory accumulation path speeds sparse matrix multiplication on wide rows, which helps algebraic multigrid setup, and rocsparse_spsort consolidates several format-specific sorting routines into one API. hipSPARSELt adopts the same compressed catalogs and reduces its disk footprint by more than 90%.

rocFFT adds just-in-time (JIT) callbacks for multi-GPU transforms through rocfft_plan_description_set_load_callback and rocfft_plan_description_set_store_callback, which let developers fuse data preparation and post-processing into multi-GPU workflows and remove redundant kernel launches. The library also switches automatically to 64-bit index arithmetic when a problem exceeds 32-bit addressing limits, including transforms with more than 2^32 elements, large strides, and high batch counts. Transforms that fit within 32-bit limits keep 32-bit indexing to avoid extra register and instruction cost.

Profiling and Telemetry

The Streaming Performance Monitor (SPM) gains a device counting mode that samples counters across the whole device while profiling is active, whether or not a kernel is in flight. Counter data was previously visible only during kernel execution. SPM supports gfx1200, gfx1201, and gfx1250, with a sample interval that is configurable in reference-clock cycles.

The ROCm Systems Profiler adds rocsys, an experimental command-line entry point that consolidates several binaries, so one command profiles an application and produces a full trace. Subcommands cover binary rewriting, call-stack sampling, instrumentation, attaching to live processes, and causal profiling. Developers can attach from the host to a running container, and AMD Infinity Storage telemetry is now exposed in AMD-SMI and the Systems Profiler, which lets teams line up I/O activity with GPU compute timelines in Perfetto and ROCm-Optiq.

AMD Simulation Toolkit

The open-source AMD Simulation Toolkit for ROCm adds four libraries for Instinct GPUs: Quadrants, a multi-platform compiler for physics simulation; Genesis, a Physical AI simulation platform; Warp, a Python framework for simulation and graphics code; and MuJoCo-Warp, a general-purpose physics engine. AMD says the toolkit lays the groundwork for its expansion into Physical AI, and it plans a follow-up post on the individual libraries.

Specification AMD ROCm 10.2.0
Serving Builds vLLM 0.32.0
SGLang 0.5.21
Python 3.14
Framework Wheels and Libraries PyTorch 2.15
JAX 0.11.2
MIGraphX 2.19
ONNX Runtime 1.29
Prebuilt AITER Kernels Instinct MI300X
Instinct MI325X
Instinct MI350X
Instinct MI355X
New Ryzen Processor Support 18 processors (gfx1103)
Storage hipFile with LVM and mdraid
PyTorch load_storage and save_storage APIs
fio v3.43
ior (main branch)
hipBLASLt Complex GEMM CGEMM and ZGEMM (gfx942, gfx950)
hipBLASLt Catalog Size About 3.8GB to about 194MB
ROCm CLI v0.3.0
Profiler rocsys (experimental)

Availability

ROCm 10.2.0 is available now, with the rocsys command marked experimental and Hyperloom support for the Radeon AI Pro R9700 in early access. AMD says the next release, ROCm 10.3.0, will arrive in the new year and may land slightly later than its usual six-week cadence.

AMD ROCm

Engage with StorageReview

Newsletter | YouTube | Podcast iTunes/Spotify | Instagram | Twitter | TikTok | RSS Feed

Harold Fritts

I have been in the tech industry since IBM created Selectric. My background, though, is writing. So I decided to get out of the pre-sales biz and return to my roots, doing a bit of writing but still being involved in technology.