Beyond base llama.cpp: what works for local LLMs on RDNA2 AMD GPUs
xpnrt · reddit · 2026-10-19
The poster notes that most AMD GPU tools in the local LLM space depend on WMMA, available only from RDNA3 onward, leaving RDNA2 users with a works-or-doesn't situation; image/video generation can sometimes bypass this with Triton, but LLMs lack such a workaround. STRATA is one of the few with an RDNA2 path (tested with a coder model), yet only for one specific large model. They ask for forks or apps that perform better on RDNA2, preferably on Windows.
More from Infra
- Hyperscaler capex shifting from chip vendors to in-house silicon by 2027, investor argues — pdamodaran · 2026-10-19
- CRWV is ~8% of Nvidia GPU sales, echoing Cisco's dot-com circular-deal trap — churchkey · 2026-10-19
- Meta details Prometheus gigawatt cluster: 100km interconnects, ~1M GPUs, 99% fewer training interruptions — ryanshrout · 2026-10-19
- Goldman upgrades CoreWeave to Buy, sees per-megawatt returns 'biased to the upside' — churchkey · 2026-10-19
- DeepSeek-V4.1-Flash quantized to 3-bit for 3x RTX PRO 6000, KLD only 0.066 nats — TheZachMueller · 2026-10-19
- SK Hynix perps rally nearly 8% in 24 hours as AI chip stocks rebound — Polymarket · 2026-10-19