Strata EXL3 AMD fork boosts decode speed and adds MiMo support
Developer nasone32 released an update to a Strata EXL3 fork for AMD, with improved HIP kernels and memory management producing a reported 10–20% average decode-speed increase, with some cases approaching 24%. The update also adds optional CPU-side expert repacking and experimental MiMo V2.6 Flash support, which the developer says ran at about 25–30 tokens per second on two Radeon RX 7900 XTX cards. Benchmarks were performed on RDNA3 hardware under Linux and ROCm, and the fork remains developed and tested primarily on that platform.
The update could make higher-quality EXL3 quantization more practical for local inference on consumer AMD systems, while the experimental MiMo support extends the fork to much larger models.
A quick update on my Strata EXL3 fork for AMD.
I've been working on the HIP kernels and memory management, and decode performance is now roughly 10-20% higher than the previous version, depending on the quant and workload. Some cases improved by almost 24%.
The other news is that I managed to get MiMo V2.6 Flash running at around 25-30 tokens/s with 2 cards.. More on that below.
First, why EXL3?
I think this is worth repeating because people tend to compare quantizations just by looking at the number of bits.
EXL3 is NOT equivalent to regular GGUF quants at the same bpw.
#Strata#EXL3#MiMo V2.6 Flash#Qwen3.8-Flash-Next#AMD#ExLlamaV3#Radeon RX 7900 XTX