Skip to content

v0.2.0

Latest
Latest

Choose a tag to compare

View all tags
@juhovainio juhovainio released this 07 Oct 16:49
v0.2.0
a5e507a
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

ROCm CLI v0.2.0 (2026-10-07)

Builds on v0.1.0 with ROCm 10.1 vLLM install support, pre-download model compatibility checks, rocm remote command surface for Tailscale based remote system management and model serving, a round of runtime-selection and diagnose correctness fixes, and WSL ROCDXG driver install support. With v0.2.0 ROCm 10.1 vLLM install has been verified end-to-end, vLLM gained OOM diagnostics and guidance, a byte-size display bug was closed across every formatter that had it, and rocm fix --json now publishes a machine-readable Doctor catalog.

New Features

  • vLLM on ROCm 10.1: rocm install vllm installs and serves working ROCm 10.1 wheels (torch 2.13.0, torchvision 0.28.0, torchaudio 2.11.0.2) from whl-next via a cp314 runtime, alongside the existing cp312 path (#471, #583)
  • vLLM out-of-memory handling: low-VRAM warnings and a --gpu-memory-utilization hint on shared/busy GPUs, a VRAM telemetry fallback via amdgpu DRM sysfs when amd-smi is unavailable on single-GPU hosts, and a new rocm diagnose/rocm fix entry for vLLM startup OOM that distinguishes a busy-GPU tenancy collision from a model that genuinely doesn't fit (#251)
  • rocm fix --json publishes the Doctor catalog as a versioned manifest (every entry, per-platform behavior, and exit code meanings), generated from the same compiled catalog the skill and CLI use, so the two can't drift (#400)
  • rocm diagnose --model <ref> answers whether a model will run on this machine from the recipe catalog and host facts, before downloading it (#407)
  • rocm remote: serve a model on a tailnet GPU machine (#333)
  • rocm diagnose --report refuses to describe hardware that isn't publicly available, instead of leaking details about unreleased/internal parts (#452)
  • rocm install sdk now installs the ROCm compiler toolchain only with --devel; a default install no longer pulls it in (#167)
  • rocm examine reports every libamd_comgr (code object manager) library found on the system, not just the first, making mismatched-runtime compile failures diagnosable (#384)
  • rocm examine --json carries the setup-recovered runtime folder through its output (#444)
  • rocm services list counts the local server records the default view hides, instead of only showing live ones (#409)
  • dashboard onboarding wizard: restored the install-folder (--prefix) choice dropped by an earlier rewrite (#441)

Bug Fixes

  • Fixed: ROCm 10.1 vLLM install resolving cleanly but failing to actually run (vllm import crash, mismatched torch/torchaudio/device-kernel pins); the install now sources the correct whl-next builds end-to-end (#583)
  • Fixed: a size one step below a unit boundary (e.g. 1,048,525 bytes) rendering as 1024.0 KiB instead of 1.0 MiB, across the disk-space, download-approval, and dashboard VRAM formatters (#503, #556)
  • Fixed: rocm diagnose claiming a fix would be auto-applied when rocm fix only reports it (#382)
  • Fixed: rocm fix fix-8-wheel-rocm offering only ROCm 6 torch indexes to ROCm 7 users (#453)
  • Fixed: rocm examine missing AMD Instinct GPUs, which enumerate under a different PCI device class than consumer/workstation cards (#434)
  • Fixed: fix-5-amdgpu-load's reported needs_reboot disagreeing with the fix catalog depending on which code path answered (#450)
  • Fixed: runtime selectors matching on substring instead of the exact key, picking the wrong runtime (#497)
  • Fixed: runtime rollback target getting lost when re-selecting the active runtime (#485)
  • Fixed: runtime path comparisons being done as text instead of by the folder they name (#491)
  • Fixed: rocm services prune sweeping a service's leftover files mid-launch instead of waiting for it to finish (#419)
  • Fixed: rocm install driver on WSL2 pointing users at a setup script that isn't shipped, instead of installing ROCDXG itself (#224)
  • Fixed: dashboard TUI command/argument previews not quoting arguments containing whitespace (#500)
  • Fixed: dashboard TUI job status glyph, label, and color drifting out of sync across renderers (#414)
  • Fixed: dashboard amd-smi safety preflight not recognizing WSL2's DXG transport (#375)
  • Fixed: rocm-doctor skill's fix-9 auto-apply flag (#149)

Breaking Changes

  • rocm install sdk no longer installs the ROCm compiler toolchain by default; pass --devel if you build GPU code (#167)

Known Issues

  • Managed Lemonade ROCm backend can hang on install, or use excessive disk space, on Windows (#392)
  • Instinct serving on RHEL 9.6 can still require manual vLLM/OpenMPI/amdsmi workarounds (#254)
  • Lemonade's llamacpp:rocm backend pin can't align to ROCm 10.1 yet; upstream (lemonade-sdk/llama.cpp) hasn't published a matching build. rocm-cli correctly reverts to the last-known-good pin, which still works for GPU inference since ROCm's sonames are unchanged across versions (#553)

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 -- v0.2.0
Assets 14
Loading