kyuz0 Strix Halo sweep: GLM-4.7-Flash-UD-Q8_K_XL vulkan_radv decode_tok_s on Radeon 8060S (gfx1151)
41.17 tok/s (tg128, batch 1)
Fields
id
benchmarks/kyuz0-strix-8060s-glm47-flash-q8kxl-vulkan-radv-tg128Metric
decode_tok_s
Value
41.17
Unit
tok/s (tg128, batch 1)
Model
GLM-4.7-Flash-UD-Q8_K_XL
Methodology
kyuz0 amd-strix-halo-toolboxes automated llama-bench sweep (results.json, generated 2026-05-16, Fedora 43 kernel 6.19.12, llama.cpp build 9187 0253fb21f). Environment: vulkan_radv on AMD Ryzen AI Max+ 395 / Radeon 8060S (gfx1151), -fa 1 -ngl 99 --mmap 0, default pp512/tg128 context. Per-request batch-1, NOT aggregate. mean +/- std over -r repeats.
Measured by
third_party
Reproducible
yes
notes
GLM-4.7-Flash ~30B MoE. 554-run machine-readable sweep; values are per-request batch-1.
Notes
GLM-4.7-Flash ~30B MoE. 554-run machine-readable sweep; values are per-request batch-1.