Skip to content

rocm-cli v0.3.0

Latest

Choose a tag to compare

@juhovainio juhovainio released this 18 Nov 16:36
v0.3.0
07f72bb

ROCm CLI v0.3.0 (2026-11-18)

ROCm CLI v0.3.0 builds on v0.2.1. It adds an AIM serving engine, an Anthropic-compatible /v1/messages endpoint, safer serving and runtime changes, and repo-native driver installs, along with a large round of WSL, Windows, dashboard, install and diagnose fixes.

Highlights

  • AIM serving engine. --engine aim serves AMD Enterprise AI AIM Docker containers alongside vLLM and Lemonade. It is opt-in
  • Anthropic-compatible endpoint. rocm serve exposes /v1/messages, including tool use and streaming
  • Follow a starting server. rocm logs --service <id> --follow streams a server's output, including download progress, until it is up
  • Safer serving and runtimes. rocm serve picks a free port without collisions, and runtime activation and rollback are transactional
  • Newest stable ROCm by default. rocm install sdk now installs the newest stable release (ROCm 10.2 today) and takes --python-version for ROCm 10 installs
  • Repo-native driver installs. rocm install driver uses the same flow on apt, dnf and zypper hosts, adds Ubuntu 26.04, and no longer uses amdgpu-install
  • Better diagnosis. rocm diagnose takes --symptom-file, estimates memory for a GGUF model, and flags an outdated Visual C++ runtime. vLLM startup failures now name the failing line
  • Runtime visibility. rocm runtimes list and rocm examine show where each runtime came from, and rocm runtimes list --json prints JSON
  • Lemonade on Linux. Lemonade is pinned to 2026.44.1 so a ROCm 10.1 SDK gets a ROCm backend, and installs check glibc and OpenSSL up front
  • Security. Endpoint API key files are owner-only on every platform, and vLLM requires the endpoint key on every server route

Breaking Changes

  • amdgpu-install is removed as an install method, with no deprecation period. rocm install driver on dnf and zypper hosts now uses the same repo-native flow as apt: it imports the AMD key, writes a package-manager repo file and installs amdgpu-dkms. The default driver source moves from a pinned 7.2.4 to latest (#181)
  • The ROCM_CLI_AMDGPU_VERSION and ROCM_CLI_AMDGPU_PACKAGE_RELEASE environment variables are no longer read. Set ROCM_CLI_AMDGPU_DRIVER_VERSION to pin a specific amdgpu release. rocm install driver refuses to plan while either old variable is still set, so an old pin is never silently turned into latest

Changes in Behaviour

  • rocm serve <model> on a host with no usable AMD GPU now succeeds when it matches an already-running managed service for the same engine and model, and prints the reuse summary. Every other case still refuses under the default gpu_required policy (#703)
  • The Dash serve wizard is now model selection plus inline Advanced settings, validated the same way as the CLI before approval (#243)
  • The dashboard assistant no longer advertises the list_skills and skill_plan tools, whose install recipes nothing could run (#882)
  • rocm install sdk without --version now installs the newest stable ROCm release instead of ROCm 7.14. Pin --version 7.14.1 for the earlier release (#1033)

Notable Fixes

  • rocm serve --engine vllm on native Windows no longer reports a launch that cannot work. It refuses up front with the reason (#837)
  • rocm uninstall removes the PATH change the installer made, and no longer empties a symlinked folder or leaves rocm.exe behind (#636, #562)
  • On WSL, rocm examine and rocm serve --gpu auto detect the GPU and librocdxg correctly (#915, #870, #1091)
  • rocm serve no longer hangs on a bad amd-smi driver state, and stopping a managed server escalates to SIGKILL when an engine ignores the polite stop (#447, #842, #1088)
  • Managed services stop reliably, and an unconfirmed stop is no longer reported as stopped (#799, #960)
  • vLLM installs on ROCm 10.x no longer fail on an unsatisfiable amd-aiter pin (#755)
  • Automation stores no longer lose records to concurrent writers or grow without bound (#738, #744)
  • The dashboard runs exactly the tool call its approval card shows (#980)
  • An unreadable config.json is no longer replaced with defaults (#541)

Known Issues

  • Instinct serving on RHEL 9.6 can still need manual vLLM, OpenMPI and amdsmi workarounds (#254)
  • ROCm SDK 10.2.0 with Lemonade 2026.44.1 (observed on Strix Halo in WSL2) copies between ordinary host memory and the GPU run at roughly 20 MiB/s, over 100× slower than on 10.1.0, so loading a large model takes many minutes. rocm serve with Lemonade stops the server at its 2-minute readiness timeout and marks the service failed. Workaround: install ROCm SDK 10.1.0 with rocm install sdk --version 10.1.0, or rocm runtimes rollback if a 10.1 runtime was the previous one, and start serving with rocm serve <model> again. (#1143)
All new features
  • --engine aim: serve AMD Enterprise AI AIM pre-built Docker containers as a third serving engine, alongside vLLM and Lemonade. It is opt-in, and the catalog gains the AIM images, which resolve through rocm serve and show in rocm model --verbose (#708, #794)
  • rocm logs --service <id> --follow (also rocm services logs <id> --follow) streams a starting server's output, including in-place download progress bars, until the server is no longer starting or running. Ctrl-C stops following and leaves the server running (#857)
  • rocm serve without --port now picks the first free loopback port in 11435..=11535. Port claims are serialized between the CLI and rocmd, so two launches can no longer take the same port; an explicit --port behaves as before (#243)
  • rocm runtimes activate and rollback, and the --activate tail of rocm update and rocm install sdk, are now transactional. They name the running servers that keep the old runtime, and a failed second write no longer leaves the config and the active-runtime marker disagreeing (#424)
  • rocm install driver plans the Ubuntu 26.04 (resolute) DKMS flow and warns about amdgpu-dkms left by a previous install, without removing anything (#778)
  • rocm examine shows the AMD display driver version, date and age on Windows, and flags drivers older than the first HIP SDK for Windows (#680)
  • rocm diagnose flags a Visual C++ runtime older than 14.40, which can crash programs such as Lemonade's llama.cpp server. rocm examine --json gains msvc_redist_version, and a new print-only fix covers it (#679)
  • rocm examine detects gfx90c (Renoir/Cezanne Vega iGPUs) instead of mapping it to gfx90X-dcgpu (#648)
  • vLLM serve summaries give out-of-memory diagnostics and guidance (#284)
  • When vLLM aborts at startup because its engine cannot import (for example a missing libtorch_cuda.so), the failure now names the failing line and the rocm diagnose --symptom '<line>' command that diagnoses it. fix-17-torch-dlpack now covers Linux and WSL (#646)
  • The dashboard shows recent serving throughput in the instance detail view, shows starting services on the Home and Serving tabs, and collects GPU and memory metrics in-process on Windows (#721, #858, #715)
  • The dashboard Diagnose screen can prepare a Doctor report: p previews what it would carry (nothing is sent), and s sends only after approval, with Deny as the default (#647)
  • The Doctor skill stops when the CLI predates the failure catalog it describes, and covers the two vLLM startup failures (out of memory, and a libtorch_cuda.so import error) (#641, #644)
  • rocm diagnose --report now says, before offering the mail, that the mail goes from your own account and carries your address, and that the report itself names no user, host or file path (#758)
  • The dashboard asks for confirmation before quitting while a model is still being served (#595)
  • rocm install prints cleaner progress and a completion banner (#207)
  • rocm serve exposes an Anthropic-compatible /v1/messages endpoint, including tool use and streaming requests (#883, #986, #987)
  • rocm install sdk --python-version X.Y chooses the Python for a ROCm 10 SDK install. An install that does not need vLLM can use 3.10 to 3.14; the default stays 3.14 (#989)
  • rocm install sdk reads the current stable and nightly TheRock hosts by default (stable.repo.amd.com, nightly.repo.amd.com) and keeps earlier pins on the old ones. Without --version it now installs the newest stable release (ROCm 10.1 today) (#1033)
  • rocm diagnose --symptom-file <path> takes the symptom text from a file or stdin, and each match says whether the symptom text contributed to it (#1048, #1094, #1051)
  • rocm diagnose estimates the context size and memory needed for a GGUF model (#826)
  • rocm fix --json describes each published exit code (#1067)
  • Lemonade installs announce themselves, show their progress, and are reported by rocm storage (#853)
  • Lemonade managed-serve failures capture the backend's identity (#912)
  • The dashboard shows one node-efficiency figure everywhere it appears, with a held flag (#1017), attributes VRAM and power to managed rocm serve services (#1026), and lists running managed services on Windows (#1002)
  • The dashboard's chat wait for a reply is two minutes by default and configurable (#1021)
  • An engine caller can withhold an inherited endpoint key file (#1034)
  • rocm runtimes list and rocm examine show where each runtime came from, and rocm runtimes list --json prints the list as JSON (#1090, #1098)
  • vLLM discovery recognizes ROCm 10.2 builds and local version tags of the form .rocmNNN (#1109)
  • The engine stdio protocol is now versioned, and an engine built for an incompatible protocol is refused with a clear message instead of failing later (#1073)
  • On Linux, rocm engines install lemonade and starting Lemonade check the host's glibc and OpenSSL first and say what falls short instead of failing at launch. ROCM_CLI_DISABLE_LEMONADE_HOST_LIBRARY_CHECK=1 turns the check into a warning for hosts that provide the libraries another way (#855)
All bug fixes
  • Fixed: rocm install sdk --dry-run downloading uv and installing a managed Python on a host without one. The preview now says a real install would prepare Python (python_note:) and changes nothing (#906)
  • Fixed: rocm uninstall --dry-run creating the folders it then listed for removal (#907)
  • Fixed: rocm uninstall leaving behind the PATH change the installer made, and on Windows leaving the running rocm.exe in the install folder. The installers now record their PATH change so uninstall can undo exactly it (#636)
  • Fixed: on WSL, rocm examine reporting wsl_rocdxg_missing and rocm serve --gpu auto refusing with no usable AMD GPU detected when librocdxg.so is reachable only through LD_LIBRARY_PATH (#915)
  • Fixed: rocm examine --json not reporting an AMD GPU on WSL when the Windows host has one (#870)
  • Fixed: rocm serve --engine vllm on native Windows printing "managed service launched" and exiting 0 while the engine died with no output. It, and rocm engines install vllm, now refuse up front and give the platform reason (#837)
  • Fixed: rocm serve on Windows reporting a managed server as started when it exited immediately. It now fails fast and shows the startup log (#243)
  • Fixed: rocm services reporting an unconfirmed stop as stopped (#799)
  • Fixed: the dashboard daemon dropping managed services when the registry could not be read, and keeping a stale Lemonade instance after its host or port changed (#927)
  • Fixed: rocm serve hanging, and surviving Ctrl-C, when amd-smi ran against a kernel driver in a bad state. The /dev/kfd pre-flight now runs at every amd-smi launch site, and an amd-smi child can no longer outlive its timeout (#447, #842)
  • Fixed: --gpu accepting an index Windows does not have (#836)
  • Fixed: rocm diagnose ranking fix-13-hip-sdk-missing first on hosts with no HIP-based setup (#676)
  • Fixed: lines printed for Windows by rocm diagnose not running in PowerShell (#677)
  • Fixed: a host ROCm version with a non-numeric build suffix, such as 7.2.4-98rc1, being silently read as 7.2.4. It is now treated as unknown (#728)
  • Fixed: with ROCM_CLI_CONFIG_DIR and ROCM_CLI_DATA_DIR set but no ROCM_CLI_CACHE_DIR, rocm storage and rocm uninstall acting on the default ~/.rocm/cache instead of <data>/cache (#834)
  • Fixed: an unknown service id surfacing the raw OS error instead of saying the service was not found (#839)
  • Fixed: automation stores (events, audit, proposals) losing records to concurrent writers (#738). A store whose torn tail could not be repaired by atomic replace, for example on a full disk or when Windows security software held the file open, could stop rocmd from starting; it is now repaired in place (#822)
  • Fixed: automation stores growing without bound. They are now capped by rotation and compaction (#744)
  • Fixed: on native Windows, a detached rocm daemon staying alive after rocm serve and rocm services stop. It now runs only while an automation needs it, and rocm services stop and the uninstall plan say what remains (#838)
  • Fixed: rocmd reporting a timeout when a child wrote more output than the pipe holds, because it did not drain the pipes until the child exited (#526)
  • Fixed: rocm engines install lemonade failing on every run when a kept ROCm alignment pin names a download that is no longer served. The packaged backend pins are now restored (#754)
  • Fixed: vLLM installs on ROCm 10.x failing with an unsatisfiable amd-aiter pin. The flash-attn and amd-aiter pins now come from what the discovered vLLM depends on (#755)
  • Fixed: /etc/os-release being parsed differently by the driver plan, rocm examine and the OpenMPI hint, so a quoted ID='ubuntu' read as an unsupported distro. It is now read by one parser, and /usr/lib/os-release is used when /etc/os-release is absent (#544, #675)
  • Fixed: on Windows, rocm diagnose --report --send and the dashboard send action never opening a mail client, because cmd split the link at & (#733)
  • Fixed: command previews that did not read back as the exact arguments that run. Arguments are now quoted for a POSIX shell, including in copyable lines (#563)
  • Fixed: CLI output such as rocm model showing an em dash as ΓÇö on Windows consoles using the OEM code page. Output text is now ASCII (#683)
  • Fixed: the dashboard Hardware tab labelling binary VRAM figures as GB instead of GiB (#699)
  • Fixed: the dashboard header warning badge (⚠ N) giving no way to read the warnings. It now opens their text (#793)
  • Fixed: the dashboard Services manager painting every service's status as healthy, including stopped and failed ones (#850)
  • Fixed: the dashboard showing an idle server's all-time TTFT and TPOT indefinitely. Values are now held, marked with *, and then cleared (#812)
  • Fixed: dashboard chat tool calls stalling other dashboard work while a slow probe ran (#801)
  • Fixed: chat with the ChatGPT sign-in silently ignoring --temperature, --top-p and --max-tokens. It now warns (#704)
  • Fixed: rocm install sdk --dry-run for a ROCm 10 SDK fetching uv; the preview no longer downloads anything (#954)
  • Fixed: the SDK's torch stack being dropped or replaced when torch alignment is disabled, and the ROCm 10 SDK's torch and torchvision not being bounded to the series vLLM pairs (#995, #1031, #1080)
  • Fixed: the installer not configuring fish when fish is not the login shell (#816)
  • Fixed: rocm version reporting the frozen 0.1.0 instead of the release tag (#818)
  • Fixed: rocm examine not reporting amdkfd_loaded and amdgpu_loaded when the drivers are built into the kernel (#694)
  • Fixed: an unreadable config.json being replaced with defaults (#541)
  • Fixed: log tails reported as empty when the log had content (#961)
  • Fixed: rocm uninstall printing a wrong dry-run footer and answering --force with --force-dev-binaries (#916)
  • Fixed: the update check preparing a managed Python (#951)
  • Fixed: a [y/N] answer that is not understood being treated silently as no; the prompt now says so and ends the line on EOF (#982)
  • Fixed: managed-service recovery from the embedded and the standalone daemon, and stopping managed services through one routine shared by rocm and rocmd, with Windows identity, liveness and tree-kill hardening (#948, #960)
  • Fixed: serving a Lemonade model registered by name on Linux, and discovering serving models in the dashboard (#963, #631)
  • Fixed: a variantless Hugging Face repo being passed to Lemonade; it is now refused up front (#919)
  • Fixed: the catalog's small assistant pointing at the wrong GGUF build for Lemonade (#1044)
  • Fixed: local Qwen3 chat answers lost to unbounded reasoning in the dashboard, and dashboard chat tool output swamping small-context models (#910, #1041)
  • Fixed: dashboard Home hero showing a -0 throughput, the two Observe heroes disagreeing on throughput, ADLX initialization on Windows, and CPU graphs not rendering continuously (#914, #1030, #1020, #856)
  • Fixed: Doctor skill and rocm fix wording in several places (#976, #1014, #1053, #1093)
  • Fixed: a ROCm 10.1 SDK getting no usable Lemonade ROCm backend. Lemonade is now pinned to 2026.44.1, which has one (#855)
  • Fixed: rocm serve not saying in its summary when the post-launch smoke test fails (#1050)
  • Fixed: a managed runtime on WSL not loading its own librocdxg, and not saying when it does not (#1091)
  • Fixed: a second amd-smi library ending up in the runtime's torch (#1064)
  • Fixed: a vLLM install being called usable before its pinned tensorizer imports (#1092)
  • Fixed: a timed-out rocm fix step not being reported as a timeout (#1062)
  • Fixed: installer download progress drawing outside the terminal (#847)
  • Fixed: running ROCm 7 and ROCm 10 runtimes side by side, and a removed rollback target staying in the active-runtime marker (#1106)
  • Fixed: the dashboard running a different tool call than the one its approval card shows. It now runs exactly the call you approved (#980)
  • Fixed: Lemonade service state being written non-atomically, so concurrent writers could corrupt it (#1045)
  • Fixed: rocmd leaving an engine child running when the supervisor could not record it (#1039)
  • Fixed: Lemonade and vLLM teardown not escalating to SIGKILL when an engine ignores the polite stop (#1088)
  • Fixed: the natural-language plan shown to you being built separately from the plan that runs, so a cloud planner's non-deterministic answer could make them differ. The plan is now built once per request (#741)
  • Fixed: the service id being dropped from inline key=value service rows (#698)

Contributors

Thanks to @jussielo-amd, @rominf, @volen-silo, @juhovainio, @tomastola, @r0x0r, @pmoutsias-amd, @johnl-amd, @siloteemu, @fredespi, @nowycondro, @michaelroy-amd and @haribabug for the changes since v0.2.1.

Installation

Install the latest stable release on Linux x86_64:

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

Install the latest stable release on Windows x86_64 from PowerShell:

$script = "$env:TEMP\install-rocm-cli.ps1"; irm https://raw.githubusercontent.com/ROCm/rocm-cli/main/install.ps1 -OutFile $script; powershell -ExecutionPolicy Bypass -File $script

Install a specific tagged release:

curl -fsSL https://raw.githubusercontent.com/ROCm/rocm-cli/main/install.sh | sh -s -- v0.3.0