Luce DFlash Support for AMD Strix Halo

The open-source project lucebox-hub has released DFlash and PFlash support for the AMD Ryzen AI MAX+ 395 (Strix Halo) iGPU. This is the result of optimizing the existing Luce DFlash stack for consumer AMD APU environments.

Benchmark results comparing performance against llama.cpp HIP on the Qwen3.6-27B (Q4_K_M) model are as follows:

• Decode: 26.85 tok/s (2.23x improvement)
• Prefill (16K context): 20.2s (3.05x improvement)

For a workload of a 16K prompt and 1K generation, total execution time was reduced from 147 seconds to 58 seconds, recording 2.5x faster end-to-end performance.

In particular, by leveraging 128 GiB of unified memory, it offers the advantage of efficiently running large-scale models (such as Qwen3.5-122B-A10B) locally that are difficult to run on consumer GPUs with the existing 24 GiB VRAM.