Releases·Tool Release
MacBook Pro M5 Max LSE LLM running an AMD Radeon AI PRO R9700 over Thunderbolt 5 in a Razer enclosure
A developer released a macOS port of the Linux amdgpu driver plus an LLM engine (LSE, or LemonSeed-Engine), enabling an AMD Radeon AI Pro R9700 in a Razer Core X V2 enclosure over Thunderbolt 5 to run LLMs on Apple Silicon Macs. The setup targets MacBook Pro, Mac mini, and Mac Studio, and the code is published across GitHub repos including the lemonade-sdk org. It matters to practitioners running local inference because it adds a non-Apple GPU path on macOS.
68 engagement1 sourceThu, Sep 24, 2026, 07:09 AM
AI Overview
- Developer Geramy ported the Linux amdgpu driver to macOS, enabling AMD Radeon AI Pro R9700 use on Thunderbolt-equipped Macs via a Razer Core X V2 eGPU
- The LSE (LemonSeed-Engine) LLM engine supports Qwen3.6 3.8B Dense and MoE models with MTP, running on Mac only through the ported amdgpu driver
- LSE's HIPC backend works best on Linux; only the Loom/HRX backend combo runs on Mac, with HRX being a new AMD system for running both kernel types on amdgpu
- The developer next targets M-series iPad support and says the feature may or may not appear in Lemonade; code is on GitHub (lemonade-sdk/mac-amdgpu, Geramy/LSE)
Entities
Related
Releases·Tool Releasefeatured
Qwen3.8-27B hits 159 tok/s on R9700 via iPad eGPU as LemonSeed Studio runs on-device inference on iPadOS
LemonSeed Studio 0.5.8 brings on-device inference to iPad via an AMD Radeon AI PRO R9700 in a Thunderbolt enclosure, embedding the upstream Linux amdgpu/amdkfd driver as a PCIDriverKit extension. Its LemonSeed Engine (LSE) records forward passes as graphs, fuses ops, generates per-GPU kernels, and uses measured speculative decoding (MTP and DFlash2) to reach 158.5 tok/s decode on Qwen3.8-27B Q4 at 131k context.
LemonSeed Studio 0.5.8 brings on-device inference to iPad via an AMD Radeon AI PRO R9700 in a Thunderbolt enclosure, embedding the upstream Linux amdgpu/amdkfd driver as a PCIDriverKit extension. Its LemonSeed Engine (LSE) records forward passes as graphs, fuses ops, generates per-GPU kernels, and uses measured speculative decoding (MTP and DFlash2) to reach 158.5 tok/s decode on Qwen3.8-27B Q4 at 131k context.
246 engagement1 sourceredditFri, Oct 9, 2026, 01:20 AM
Releases·Tool Releasefeatured
Unsloth announces official AMD GPU support for local LLM fine-tuning and inference
Unsloth now officially supports AMD hardware for local inference, fine-tuning, reinforcement learning, and deployment. The open-source tool works on Windows, Linux, and WSL with Radeon RX 9000/7000 series, Instinct MI350/MI300 GPUs, Strix Halo/Ryzen AI Max systems, and AMD CPUs for GPU-free inference. It claims up to 70% less VRAM for training and 80% less for reinforcement learning.
Unsloth now officially supports AMD hardware for local inference, fine-tuning, reinforcement learning, and deployment. The open-source tool works on Windows, Linux, and WSL with Radeon RX 9000/7000 series, Instinct MI350/MI300 GPUs, Strix Halo/Ryzen AI Max systems, and AMD CPUs for GPU-free inference. It claims up to 70% less VRAM for training and 80% less for reinforcement learning.
1101.2k engagement3 sourcesreddit, github, youtubeMon, Jul 20, 2026, 02:48 PM
Developments·Hardware
User questions AMD Radeon R9700 dual-GPU setup for local LLMs vs NVIDIA CUDA
A user built a local AI server with two Radeon AI PRO R9700 GPUs (64 GB total VRAM) and asks whether AMD/ROCm is mature enough for running LLMs compared to NVIDIA/CUDA, citing concerns about compatibility, performance, and model quality.
A user built a local AI server with two Radeon AI PRO R9700 GPUs (64 GB total VRAM) and asks whether AMD/ROCm is mature enough for running LLMs compared to NVIDIA/CUDA, citing concerns about compatibility, performance, and model quality.
51 engagement1 sourceredditThu, Jul 30, 2026, 07:35 AM
Releases·Hardware Release
BaseRT inference runtime exploits Apple M5 Neural Accelerators for LLM throughput gains
Apple's M5 chip introduces dedicated Neural Accelerators on each GPU core, exposed via Metal 4 tensor API. The BaseRT runtime, a native Metal inference engine, uses hand-written tensor-core kernels to outperform llama.cpp and MLX on Apple hardware. The paper details dense and MoE GEMM and flash-attention prefill kernels.
Apple's M5 chip introduces dedicated Neural Accelerators on each GPU core, exposed via Metal 4 tensor API. The BaseRT runtime, a native Metal inference engine, uses hand-written tensor-core kernels to outperform llama.cpp and MLX on Apple hardware. The paper details dense and MoE GEMM and flash-attention prefill kernels.
0 engagement1 sourcearxivTue, Jul 21, 2026, 06:42 AM
Releases·Tool Releasefeatured
Open-source kernel hits 78,498 output tok/s for Qwen3.6-35B-A3B on 8 AMD MI350X GPUs
A developer team optimized AMD's software stack down to the kernel level for Qwen3.6-35B-A3B on MI350X GPUs, achieving 78,498 output tokens per second across 8 GPUs. The work aims to close the software gap with NVIDIA's CUDA ecosystem, leveraging MI350X's raw compute advantage over B200.
A developer team optimized AMD's software stack down to the kernel level for Qwen3.6-35B-A3B on MI350X GPUs, achieving 78,498 output tokens per second across 8 GPUs. The work aims to close the software gap with NVIDIA's CUDA ecosystem, leveraging MI350X's raw compute advantage over B200.
80 engagement1 sourceredditWed, Aug 26, 2026, 04:12 AM