lhl/strix-halo-testing llm-bench — Vulkan AMDVLK long-context sweep on Strix Halo (gpt-oss-120b)

llama-bench b6608 (commit 1d108eef7), Vulkan AMDVLK 2025.Q3.3 (2.0.355) on a GMKtec EVO-X2 Strix Halo, Ryzen AI MAX+ 395, 128 GB RAM allocated 96 GB VRAM, Linux 6.16.4, performance profile, power-profiles-daemon disabled. Model unsloth gpt-oss-120b-F16.gguf (MXFP4 MoE + Q8_0 attention, 65.4 GB). Flags: -fa 1 --no-mmap -ngl 999, prompt 2048, gen 32, context depths 0 and 32768. Per-request batch-1 pp/tg, NOT aggregate. Same run family as the short-context lhl rows already in the repo, but this file is the long-context depth sweep.