llama.cpp issue #28863: ROCm/gfx1100 Qwen3.8-27B Q4_K_XL batch-1 tensor-split decode on 2x RX 7900 XTX

GitHub issue #28863 in ggml-org/llama.cpp about ROCm/gfx1100 tensor-split decode performance. Opened 2026-09-13 by LithiumDevourer. Qwen3.8-27B Q4_K_XL batch-1 tensor-split decode reaches only 1.31x one card. Verified 2026-10-10: Read issue 'Research: ROCm/gfx1100: Qwen3.8-27B Q4_K_XL batch-1 tensor-split decode reaches only 1.31x one card, with per-card DRAM utilisation falling from 66% to 43%' (https://github.com/ggml-org/llama.cpp/issues/28863); confirms cited findings by LithiumDevourer including 1.31x scaling and per-card DRAM drop from 66% to 43%.