Lemonade 11.9 Local AI Server Released With Super Exciting AMD ROCm HRX Backend

Lemonade 11.9 is out today as the newest feature release to this AMD-backed, open-source local AI server solution across Linux, Windows, and macOS. Lemonade has long been focused on offering "100% free and private" AI use with local hardware whether it be GPUs, CPUs, or NPUs. With Lemonade 11.9's release today it's very interesting for having experimental ROCm HRX back-end support with Llama.cpp. HRX is the new exciting thing to watch out for on the AMD ROCm compute landscape. Exciting me the most with Lemonade 11.9 is that it has a Llama.cpp HRX back-end working with the AMD HRX Runtime and initially for Radeon RX 7900 series RDNA3 and Strix Halo APUs. HRX is a new development in the ROCm ecosystem and part of their Loom/Hyperloom efforts announced back at the AMD Advancing AI event in San Francisco. Hyperloom was announced at AAI as part of their agentic workload optimization efforts. HRX is part of their Loom compiler and IR stack as an alternative to their long-used LLVM IR and getting the generated code quicker into the optimized AMDGPU assembly. Not only that but it also is a common component for helping get to high performance integration across all AMD GPUs / NPUs / CPUs. The hrx-system GitHub repository explains: "The HRX System is a collection of minimal runtime components providing an alternative implementation of HIP as it presently ships within ROCm. It provides a common substrate for low latency, high performance integration with AMD's GPU, NPU, and CPU products." This is quite interesting and reminds me of the long-talked-about AMD Unified AI Software stack and in the past talking about MLIR and SPIR-V usage as their preferred common intermediate representation language. Now Loom IR comes into the equation as their custom and more catered IR for AMD hardware. More details on HRX in the context of Llama.cpp was brought up in the past two weeks via this GGML discussion thread started by AMD engineer Stella Laurenzo. Laurenzo explained in the initial post: "At AMD, we are actively optimizing AMD GPU performance in llama.cpp under the existing Vulkan and HIP (CUDA-derived) backends. However, as ROCm matures on client systems, we have repeatedly run into the costs of not having AMD native backends and optimized client libraries that can be independently optimized for the best possible performance of the GPUs and NPUs on our silicon. In addition, as ROCm was originally developed primarily to service the datacenter market, it often faces headwinds when trying to integrate and deploy to client operating systems in a way that best meets the needs there. As a result, we have been working on a lighter, more focused subset of ROCm that is optimized for these use cases. We call this HRX, and we have been co-developing it with the needs of a dedicated llama.cpp backend for a couple of months, for eventual inclusion into ROCm proper as a focused, redistributable set of libraries. While the work is still early, we wanted to discuss with the community the possibility of finishing this work and landing it as a new ggml-hrx backend in llama.cpp. While we have done a lot of experimental work to optimize this backend for a wide swath of models and use cases, we decided that for an initial debut and discussion, we would focus the contribution on a minimal kernel library sufficient to run the unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF Q4_K_M model performantly on RDNA3 (W7900) and RDNA3.5 (Strix Halo). This RFC PR carries that minimum set, and we will continue to develop the branch for wide model, operating system, and system coverage." The post goes on to note with HRX the possibility of 30~50% tok/s uplift on prefill compared to Llama.cpp/GGML's existing Vulkan and HIP backends. There's also parity to 10% token per second uplift on non-MTP decode tasks. The HRX stack is also simpler for development. ROCm HRX is extremely exciting from a technical perspective albeit something that ideally would have come out years ago rather than only now in experimental state in H2'2026. With today's Lemonade 11.9 release is the experimental llamacpp-hrx back-end working for HRX on AMD GFX1100 and GFX1151 GPU targets under Linux. Hopefully the GPU support will be quickly expanded moving forward and I am eager to test this myself. Lemonade 11.9 also brings Qwen3-Next support to its pinned Llama.cpp back-end and other improvements.