AMD Ryzen AI Developer Platform — Release Notes

Current Stack Versions

The versions provided as part of this update have been validated by AMD and partners to ensure the current stack works seamlessly.

Component	Current Version	Details
ROCm	10.0	ROCm release: https://rocm.docs.amd.com/en/docs-10.0.0/release/changelog.html
Developer Center	1.4.0	Developer Center release
Kernel	7.2.7-1+rex1	Linux kernel release: https://cdn.kernel.org/pub/linux/kernel/v7.x/ChangeLog-7.2.7
BIOS	4.00	System BIOS release
ComfyUI	0.34.2	Upgraded from 0.21.1 https://github.com/Comfy-Org/ComfyUI/releases#release-v0.34.0
fwupd	2.1.7-4~bpo13+1+rex2	Upgraded from 2.1.4 https://github.com/fwupd/fwupd/releases/tag/2.1.7
ggml	0.22.0-1+rex1	Upgraded from 0.13.1 https://github.com/ggml-org/ggml/releases/tag/v0.22.0
llama.cpp	0.2.0+dfsg1-1+rex2	Upgraded from 9413+dfsg-1 https://github.com/ggml-org/llama.cpp/releases/tag/v0.2.0
Lemonade	11.8.1-1+rex1	Upgraded from 11.5.0-1 https://github.com/lemonade-sdk/lemonade/releases/tag/v11.8.1
Node.js	20.19.2+dfsg-1+deb13u3	No change
Python3	3.13.5-1	No change
PyTorch	2.13.0+dfsg-0+rex2	Upgraded from 2.10.0-0 https://github.com/pytorch/pytorch/releases/tag/v2.13.0
TorchAudio	2.11.0.2-1+rex1	Upgraded from 2.10.0-0 https://github.com/pytorch/audio/releases/tag/v2.11.0
TorchVision	0.28.0-0+rex1	Upgraded from 0.25.0-0 https://github.com/pytorch/vision/releases/tag/v0.28.0
vLLM	0.27.1	Upgraded from 0.22.1 https://github.com/vllm-project/vllm/releases/tag/v0.27.1
amd-debug-tools	0.2.20-1+rex1	No change

New Features

• ROCm 10.0
• This release focuses on AI inference and developer tooling. Highlights include expanded framework support for AI inference, new HIP APIs and performance improvements, ROCprofiler-SDK adoption across AI profiling workflows, and updates to math, sparse, and communication libraries.
• https://rocm.docs.amd.com/en/docs-10.0.0/about/release-notes.html#release-highlights
• Developer Center v1.4.0
• Developer Center now has a Developer Platform Security section added to the Manage tab (when security updates are available) with a dedicated security update channel. This ensures end users have a direct channel to receive CVE fixes from Debian.
  
  The app checks security updates automatically in the background once a week, so no manual action is needed to stay informed. The update automatically applies after 5 days and requires a reboot. The user can manually update as well by clicking the update button. Clicking this button will not update your system configuration.
• Lemonade 11.8.1
• A new experimental TheNoise ROCm image-generation backend serves Anima and Krea-2
• A new experimental ds4 (DwarfStar) backend serves DeepSeek V4 Flash
• The Muse-Glimmer-30B-GGUF model joins the built-in llama.cpp catalog with draft decoding, vision, and tool-calling support
• The catalog adds the RPG-HaloTales-V2 chat model, Flux-2-Klein-4B/9B image models, and new OpenMOSS voice-design and sound-effect speech models
• https://github.com/lemonade-sdk/lemonade/releases#release-v11.8.1
• vLLM 0.27.1
• Unified Memory & VRAM: Memory handling for shared RAM was stabilized, eliminating memory detection bugs and allowing larger context sizes
• Kernel & Architecture Support: ROCm and Triton support for the gfx1151 architecture matured, fixing quantization issues and removing the need for custom build patches
• Engine Efficiency: The shift to Model Runner V2 reduced host CPU overhead and bus contention, delivering smoother and faster token generation
• Clustering & Interconnects: Inter-node communication was improved, making multi-device tensor parallelism on Strix Halo setups more reliable and lower latency
• https://github.com/vllm-project/vllm/releases#release-v0.27.1
• PyTorch 2.13.0
• Add support of gain expandable memory segments, rocSHMEM symmetric memory collectives, and FlexAttention pipelining
• Improved Performance with AOTriton
• https://github.com/pytorch/pytorch/releases/tag/v2.13.0
• Power Management
• System suspend is enabled by default with a 5 minute delay (screen blank) and 15 minute delay (suspend). While this feature is enabled, there is the suspend inhibition feature (available via Developer Center, enabled by default) where suspend can be inhibited if this system is being used headless or a GPU intensive workload is running.

Bug Fixes

• drm/amd/display: Set gpuvm min page size to 4K on dcn35/36
• Fix for random DMA corruption
• Fix issues with dGPU in eGPU enclosure
• Add TLB flush after MES queue eviction
• Fix handling of GPU oversubscription
• Plymouth shows duplicate logos on tiled displays during boot and shutdown
• Ensure package autoremoval on update

Security Fixes

• Kernel CVE fixes included (non-exhaustive list)
• CVE-2026-23288
• CVE-2026-23341
• CVE-2026-23424
• CVE-2026-43446
• CVE-2026-63873
• CVE-2026-68380
• CVE-2026-72090
• CVE-2026-72328
• CVE-2026-72332
• CVE-2026-89978
• CVE-2026-98144
• CVE-2026-98145
• libdrm
• amdgpu: reject out-of-range ip_instance in CS and semaphore APIs
• xf86drmMode: avoid double-free of connector arrays on hot-plug retry
• amdgpu: clamp num_shader_engines loop to array size
• amdgpu: guard against overflow in handle_table_insert
• amdgpu: fix self-deadlock in amdgpu_bo_import error path
• amdgpu: bound alloca sizes in CS submission paths
• xf86drm: fix integer overflow in allocations
• amdgpu: check drmGetVersion result before dereferencing
• amdgpu: handle malloc failure in split_env_var
• xf86drm: guard against negative read() return in drmOpenByName
• GNOME Display Manager (GDM)
• Fix many security issues identified by Mythos
• Developer Center (non-exhaustive list)
• Fix sudoers wildcard rules allowing argument injection

• The Developer Center can sometimes seem to take longer to update itself or appear stuck. If this happens - use #sudo systemctl restart halo-lp

• Latest Release Notes
• Release Notes as of September 3, 2026
• Release Notes as of June 22, 2026