Skip to content

ROCm 10.0.0 Release Notes: From 7.14.0 to ROCm.AI

2026-08-26 · ROCm Core SDK 10.0.0 · Official docs (latest) · Official Release Notes

Back to Environment · 中文

ROCm 10.0.0 is the first major version jump after the 7.x series. 7.14.0 made TheRock the production build and release foundation. 10.0.0 keeps that foundation and, for the first time, folds install, coding, and performance work into an AI-assistant-native workflow called ROCm.AI.

For hello-rocm readers, the version number matters less than these three new entry points:

ComponentIn one sentenceWhen you use it
AMD SkillsOfficial AMD optimization knowledge packaged as Skills for Claude, Cursor, and CodexYou are writing ROCm / HIP / vLLM code with an AI assistant and want AMD best practices by default
HyperloomOpen-source auto-optimization engine: profile, find bottlenecks, rewrite kernels, tune parameters, then validateInference already runs and you want more throughput
ROCm CLIInstall, verify, deploy, and manage with one command surfaceYou do not want to hand-copy pip, driver, and image commands

This page explains what changed from 7.14.0 to 10.0.0, then unpacks those three entry points. For install steps, see 00-Environment.


1. 7.14.0 → 10.0.0 at a glance

AreaROCm 7.14.0 (2026-07-15)ROCm 10.0.0 (2026-08-26)
PositioningTheRock becomes the build / release foundationDecade milestone; ROCm.AI lands on top of TheRock
Developer experienceDocs + wheels + Docker, assembled by handAMD Skills + Hyperloom + ROCm CLI
pip indexhttps://repo.amd.com/rocm/whl-multi-arch/https://stable.repo.amd.com/rocm/whl-next/
System packagesDistributed from the older repo.amd.com layoutConsolidated at stable.repo.amd.com
PyTorch2.12.02.13.0 (also 2.12.0 / 2.11.0)
torchvision / torchaudio0.27.0 / 2.11.00.28.0 / 2.11.0.2
JAX0.10.00.11.0 (also 0.10.2 / 0.10.0)
vLLM0.23.00.27.0
SGLang0.5.130.5.15
TensorFlow2.212.21 (also 2.20 / 2.19.1)
MIGraphX / ONNX Runtime2.16 / 1.23.2 generation2.17 / 1.29.0
Windows driverAdrenalin 26.5.1Adrenalin 26.8.1
Linux driveramdgpu 31.30.0 generationamdgpu 31.50.0
New hardwaregfx1153 (Ryzen AI 7 445 / AI 5 435, etc.)Radeon RX 9050 / 9050 4GB (gfx1200)
Windows SDKHIP SDK on a separate track from LinuxHIP SDK retired; Windows and Linux share the Core SDK
ASAN packagesNot shipped in 7.14.0Shipped alongside standard packages
Recommended installuv pip + official imagesuv pip still works; new path is rocm install sdk

Treat ROCm 10.0.0 docs as the canonical entry. Compatibility is defined by the Compatibility Matrix.


2. Why the jump from 7 to 10

7.14.0 answered “how do we split ROCm, install only what we need, and ship Windows / Linux from one TheRock pipeline.” 10.0.0 answers the next layer:

  • After install, developers still had to remember gfx extras, pick images, match drivers, and read optimization docs by hand.
  • AI coding assistants default to CUDA / NVIDIA habits and do not automatically take AMD kernel, scheduling, or quantization paths.
  • Inference optimization was still “human reads a profile → human edits a kernel → human re-benchmarks,” on a week-scale loop.

So 10.0.0 is not just another library bump. It moves the platform from shipping libraries to shipping workflows. The official blog puts it directly: older ROCm releases mainly gave better primitives; ROCm.AI covers the install, validate, serve, and optimize loop around those primitives. ROCm 10.0 official write-up

TheRock is not replaced. The 10.0.0 Core SDK, framework wheels, and validated images all come from the same TheRock pipeline, with minor releases about every six weeks.


3. Focus 1: AMD Skills

3.1 What it actually is

AMD Skills is not another documentation site. It is a catalog of directories written to the Agent Skills standard: each skill has a SKILL.md, plus optional scripts and references. Claude Code, Cursor, Codex, and Gemini CLI load the matching skill when the task description fits.

Docs describe every API option. A skill encodes the opinionated path an AMD engineer would take — which image, which gfx, which environment variables, and what to check before serving. When you write ROCm code with an assistant, it follows those validated paths instead of improvising from CUDA muscle memory.

Install:

bash
npx skills add amd/skills

Pin a skill and an agent:

bash
npx skills add amd/skills --skill local-ai-use --agent cursor
npx skills add amd/skills --list

You can also copy folders into ~/.cursor/skills/, ~/.claude/skills/, or $HOME/.agents/skills.

3.2 How the catalog is split

The official catalog has three layers that match the two hello-rocm paths most readers care about: local models on a Ryzen AI PC, and production inference on Instinct.

Client-native (Ryzen AI / local AI PC)

SkillWhat it does
local-ai-useRoute image generation, TTS, and STT through a local server to cut cloud token cost
local-ai-app-integrationAdd offline, privacy, and local inference to an existing cloud LLM app

Cross-stack (client to cloud)

SkillWhat it does
rocm-doctorDiagnose ROCm / HIP / PyTorch / llama.cpp against known misconfigurations; thin driver over rocm examine / diagnose / fix (planned)
lemonade-router-builderRoute requests through Lemonade by content, sensitivity, or capability
hrr-replay-analysisReplay and analyze GPU workloads with HIP Record and Replay (planned)

Server-native (Instinct / EPYC)

SkillWhat it does
serving-llms-on-instinctDetect hardware, check model fit, apply a vLLM recipe, launch a benchmarked endpoint
serving-llms-on-epycServe LLMs on EPYC with zentorch / vLLM
hyperloom-workload-optimizerInstall Hyperloom and autonomously optimize LLM inference throughput on Instinct
magpie-kernel-evaluatorEvaluate kernel correctness and performance against vLLM / SGLang
tracelens-analysis-orchestratorSplit a PyTorch profile with TraceLens and write a prioritized report

3.3 How this relates to the hello-rocm Skill

This repo’s src/hello-rocm-skill answers “which tutorial should I read, how do I deploy, how do I debug.” AMD Skills answers “what is the official AMD-validated path.” Install both:

  • “Which hello-rocm chapter should I start with, and which gfx extra do I use?” → hello-rocm Skill
  • “Stand up vLLM on MI300X the official way / squeeze throughput with Hyperloom” → AMD Skills

4. Focus 2: Hyperloom

4.1 The problem it closes

Hyperloom is an open-source agentic optimizer (source: AMD-AGI/Hyperloom). It is not another manual profiler GUI. It closes the loop of “read a trace → find the bottleneck → rewrite a kernel or retune → validate correctness and performance.”

The official loop is:

text
Profile → Analyze → Plan → Optimize → Validate → repeat

You no longer drive every cycle. Give it a model, framework, GPU, parallelism, and sequence lengths; it profiles, proposes, edits, benchmarks, and checks correctness. AMD’s claim is that week-scale manual optimization can compress into hours, while covering search space a human would skip under time pressure.

4.2 The parts underneath

ComponentRole
TraceLens-AgentMark bottlenecks from a profile automatically
MagpieEvaluate / compare kernels and benchmark vLLM or SGLang
IntelliKitConversational profiling
GEAKMulti-agent kernel rewrites across Triton, HIP, CK, FlyDSL, and TileLang
ArborSelf-evolving search over the optimization space
AgentKernelArenaA/B-test optimization agents on standard tasks

The public 10.0.0 surface targets MI300X / MI325X / MI355X, with vLLM and SGLang. On Radeon / Ryzen, hello-rocm still starts with “get deploy working.” Hyperloom is for readers who already have Instinct hardware and need production throughput.

4.3 How to attach it to your workflow

Two paths:

  1. Via AMD Skills: install hyperloom-workload-optimizer and tell the assistant to raise inference throughput on Instinct with Hyperloom.
  2. Via the official CLI: start a session from Run a Hyperloom optimization, passing framework, model path, TP / CONC / ISL / OSL, and a target gain.

Hyperloom does not replace the rocprofiler-sdk work from 7.14.0. That release made rocprofiler-sdk the PyTorch Profiler backend. 10.0.0 adds rocSHMEM / hipFile / OpenMP tracing, per-node HIP Graph attribution, and Roofline on gfx1150 / 1151 / 1152. Hyperloom consumes those traces, then takes the extra step of changing the code.


5. Focus 3: ROCm CLI

5.1 The one-command layer

ROCm CLI is the command-line face of ROCm.AI. It ships as a single prebuilt binary for Linux and Windows x86_64 and does not require an existing ROCm, Python, or Rust install. It is a Technology Preview: commands and UI can still change. It can already manage runtimes from 7.13 onward; official ROCm 10 support is listed as coming next.

Linux / WSL:

bash
curl -fsSL https://raw.githubusercontent.com/ROCm/rocm-cli/main/install.sh | sh

Windows PowerShell:

powershell
irm https://raw.githubusercontent.com/ROCm/rocm-cli/main/install.ps1 | iex

Day-to-day commands:

bash
rocm                 # launcher: setup / serve / diagnose / chat / dashboard
rocm examine         # GPU, driver, ROCm, and engine readiness
rocm install sdk     # TheRock wheels into a CLI-managed Python environment
rocm install driver  # amdgpu driver on Linux
rocm serve qwen      # local OpenAI-compatible model server
rocm dash            # full-screen telemetry (Linux / WSL)

The 7.14.0 path was “create a venv, fill extras, match image tags.” 10.0.0 adds a managed path: rocm install sdk fetches TheRock wheels and a matching PyTorch, rocm examine tells you whether the gap is driver, permissions, or runtime, and rocm serve picks an engine from the GPU — Lemonade (GGUF) on Ryzen / Radeon, vLLM (safetensors) on Instinct.

5.2 What you get beyond another install script

  • Side-by-side runtimes: rocm runtimes activate / rollback so one machine can hold multiple ROCm installs and roll back an upgrade.
  • Engine adapters: Lemonade on client GPUs, vLLM on Instinct. Windows currently has the CLI and Lemonade, not the live dashboard or vLLM.
  • A hook for agents: rocm-doctor is a thin layer over examine / diagnose / fix. Skills and the CLI share one state machine.
  • Isolation from legacy stacks: an existing 7.14.0 /opt/rocm or hand-built .venv is reported as legacy_rocm_status: detected_unmanaged. rocm install sdk creates a managed runtime next to it instead of overwriting it.

Minimum Linux is Ubuntu 24.04 (glibc 2.38+). Ubuntu 22.04 cannot run the prebuilt Lemonade engine. There is no official macOS package.

hello-rocm still documents uv pip as the reproducible baseline so wheel versions stay explicit. On a new machine, try ROCm CLI first if you just want to verify the GPU, then cross-check versions against this page and the environment baseline.


6. What else changed in the stack

6.1 Install and distribution

  • GPU software starts consolidating on stable.repo.amd.com: ROCm packages, the amdgpu driver, and public tools share one layout, with multiple versions and architectures side by side.
  • The pip index moves from repo.amd.com/rocm/whl-multi-arch/ to stable.repo.amd.com/rocm/whl-next/. The [device-gfxXXXX] extras stay the same.
  • Installed RPM / DEB / runfile packages now embed RPATH instead of RUNPATH, so a ROCm binary prefers its own library tree and is less likely to be hijacked by LD_LIBRARY_PATH on a multi-version machine. Tarballs still use RUNPATH for compose-your-own environments.
  • On Windows, the HIP SDK is retired. Windows and Linux share the Core SDK and release cadence. Windows ships as a tarball today; a native installer is planned later in 2026.

6.2 Frameworks and images

This project’s pip baseline follows the official recommendation:

bash
uv pip install --index-url https://stable.repo.amd.com/rocm/whl-next/ \
  "torch[device-gfx1151]==2.13.0+rocm10.0.0" \
  "torchvision[device-gfx1151]==0.28.0+rocm10.0.0" \
  "torchaudio==2.11.0.2+rocm10.0.0"

Official validated vLLM image (the image bundles PyTorch 2.12.0 — do not mix that with the pip 2.13.0 line):

bash
docker pull rocm/vllm:rocm10.0.0_ubuntu24.04_py3.14_pytorch_2.12.0_vllm_0.27.0

10.0.0 also adds Unsloth local fine-tuning on Ryzen AI MAX, plus ComfyUI tuning for Wan2.2, FLUX.2 KLEIN, SD 3.5, and related models. Those will land in Fine-tune / practice chapters later and do not block the 01-Deploy environment bump.

6.3 Communication, math, and tools

  • RCCL / rocSHMEM: the largest single investment in this release. RCCL’s upstream NCCL merge advances to 2.30.4, with symmetric memory, GPU-initiated networking (GIN), and Python APIs. rocSHMEM continues closing the NVSHMEM 3.6.5 API gap.
  • hipBLASLt: tune GEMM kernel selection locally for your problem shapes; weights never leave the machine.
  • ROCm Optiq 1.0 GA: Systems Profiler timelines and Compute Profiler analysis in one visualization environment.
  • ASAN packages: the Address Sanitizer builds that 7.14.0 omitted now install like any other ROCm package.

7. Upgrade guidance for hello-rocm readers

  1. New machine: follow the environment baseline with uv pip and ROCm 10.0.0. If you only want to verify the GPU first, use rocm examine / rocm install sdk.
  2. Already running 7.14.0: do not rebuild just for the version number. Upgrade when you need the new frameworks (PyTorch 2.13 / vLLM 0.27) or ROCm.AI. Remove old HIP SDK / old wheels first, and move Windows drivers to Adrenalin 26.8.1.
  3. Writing code with an assistant: keep the hello-rocm Skill and add amd/skills. They can coexist.
  4. Instinct and throughput work: treat Hyperloom as the headline of 10.0.0, not another handwritten kernel-tuning note.
  5. Local Radeon / Ryzen deploy: environment commands now target 10.0.0. LM Studio / Ollama / llama.cpp steps stay mostly the same; the ROCm and driver versions underneath change.

ResourceLink
ROCm 10.0.0 docs homehttps://rocm.docs.amd.com/en/latest/
Core SDK Release Noteshttps://rocm.docs.amd.com/en/latest/about/release-notes.html
Install selectorhttps://rocm.docs.amd.com/en/latest/install/rocm.html
Compatibility matrixhttps://rocm.docs.amd.com/en/latest/compatibility/compatibility-matrix.html
TheRock transition guidehttps://rocm.docs.amd.com/en/latest/about/transition-guide-TheRock.html
Official bloghttps://rocm.blogs.amd.com/ecosystems-and-partners/rocm-x-blog/README.html
ROCm.AI announcementhttps://newsroom.amd.com/news/rocm-10-software-ai-native-developer-experiences/
AMD Skillshttps://github.com/amd/skills
ROCm CLIhttps://github.com/ROCm/rocm-cli
Hyperloom docshttps://rocm.docs.amd.com/projects/hyperloom/en/latest/
Hyperloom sourcehttps://github.com/AMD-AGI/Hyperloom