What does it do?
This repository is a super-repo that consolidates multiple ROCm systems projects into one place to streamline development, continuous integration (CI), and integration; its first set of projects focuses on requirements for building PyTorch. [1] It organizes each ROCm systems project under projects/, where each folder corresponds to a project previously maintained in a standalone GitHub repository and released as distinct packages. [2] Its stated goals include enabling unified build and test workflows across ROCm libraries and facilitating shared tooling and CI. [3]
Activity, over time
Daily observations from the public event archive.
Attention
Daily stars & forks
Star accumulation
Cumulative stars
Pull requests
Daily opens & merges
Conversations
Daily issues & comments
Community
Daily accounts & reviews
Releases
Daily releases
Trending history
Daily top 10 · last 90 days · UTC
Latest stargazers
The last 10 distinct accounts we observed this year.
Most active participants
PR, issue, comment & review events in the past year.
README badges
Embed a badge in your README or website.
Mentions
@AIatAMD PR 3 out 3 companion. ROCm/hipblaslt was reading/decoding full 239MB elf kernel slice for a 3.5MB metadata region. All 3 Prs combined will reduce cold gemm/linear load by up to 600ms. 1.38s load is now ~0.8s. rocm-systems#13712 t.co
@AIatAMD PR moved to rocm-systems#13718
Native Windows running llama.cpp on AMD HRX backend on RDNA3.5 gfx1152 iGPU. I was able to build and run llama.cpp on the HRX backend and I am a little disappointed. AMD should stop gatekeeping features. The project currently only has ML kernels for full f16 KV cache types. It's gated behind gfx1100 and gfx1151 arches, but runs fine on any gfx11xx GPU because the arch is lowered to gfx11-generic and the HRX LOOM JIT compiler optimizes the kernels for the appropriate GPU. It is also gated behind the SVM system requirement which is unnecessary because the functionality is not used at all here. This might be an issue with HRX itself and the ROCR HSA runtime. Not sure how they preach HRX for consumers but needlessly indirectly gate it behind Linux. I am happy to report that the HRX backend beats Vulkan by over 70% on prefill and matches it on decode on my system. But it uses more VRAM under use. My ROCm/HIP backend just pooped itself after the ROCm10 update. It simply refuses to produce any output when running. So I cannot compare it here. I am slightly more impressed with the HRX backend itself compared to the llama.cpp port of it. I had to gently patch the three related repos to get them to build cleanly on Windows without them producing false positive errors. I can share git diffs or a link to my GitHub page if anyone is interested. I have not published them to my GitHub page yet because I didn't think anyone is interested lol. TLDR: The AMD HRX backend is preached to be for consumers but is still effectively gatekept behind Linux and a few select GPUs just like early ROCm. None of this gatekeeping is necessary as I have demonstrated. System Specs: OS: Windows 11 ROCR HSA Runtime 1.1 Ext. 1.32 HRX (HIP Runtime Extended) v0.1.0 Repo: ROCm/rocm-systems Branch: develop Repo: ROCm/hrx-system Branch: main Repo: AMD-Ecosystem/llama.cpp Branch: hrx-graph-develop-v2