kyuz0 Strix Halo 8060S gemma-4-26B-A4B-it-UD-Q8_K_XL Vulkan AMDVLK 64k pp2048
53.86 tok/s
Fields
id
benchmarks/kyuz0-8060s-gemma-4-26b-a4b-it-ud-q8-k-xl-amdvlk-64k-pp2048Metric
prefill_tok_s
Value
53.86
Unit
tok/s
Model
gemma-4-26B-A4B-it-UD-Q8_K_XL
Methodology
kyuz0 amd-strix-halo-toolboxes llama-bench sweep (results.json generated 2026-05-18, Fedora 43 kernel 6.19.12-200, llama.cpp build 0253fb21f / b9187). Vulkan AMDVLK, -fa 1 -ngl 99 --mmap 0. Long-context cell: context depth 65536 tokens, prompt processing of 2048 tokens. ONE request, batch size 1, so the number is per-request, not aggregate across concurrent requests. Mean of the repeated runs; std 0.04 tok/s. Model file 25.94 GiB, 25.23 B params, quant Q8_K_XL.
Measured by
third_party
Reproducible
yes
notes
Not in the earlier kyuz0 filing, which covered only the default-context pp512/tg128 cells. Same results.json, the longctx32768 and longctx65536 rows.
Notes
Not in the earlier kyuz0 filing, which covered only the default-context pp512/tg128 cells. Same results.json, the longctx32768 and longctx65536 rows.