Elpis repo brings speculative decoding speedup for Qwen3.8-27B on RX 7900 XTX
darc_knigh · reddit · 2026-11-01
A Reddit user shared the elpis-on-7900xtx GitHub repo: a recipe for running Qwen3.8-27B faster on AMD RX 7900 XTX using EXL3 quantization plus DFlash2 speculative decoding on exllamav3-rocm, with patches and benchmarks included.
More from Infra
- Microsoft Sweden data center site accused of withholding €900K wages for 78 workers — emmanuelvivier · 2026-11-02
- Probe exposes withheld wages, electrocutions at Europe's data centres — emmanuelvivier · 2026-11-02
- Back-of-envelope: machine $/FLOP just dropped below the human brain after 70 years — scychan_brains · 2026-11-02
- Judge lets NAACP lawsuit over xAI's Colossus turbine pollution proceed — emmanuelvivier · 2026-11-02
- Alleged Rebrand: OpenAI Wants to Call Its Data Centers "AI Compute Campuses" — Drastic_Conclusions · 2026-11-02
- Krea2 crawls at 20-40s per iteration on AMD RX 9060XT under ROCm — FoggyFreeze · 2026-11-02