Home › Evidence › Records › cfg-0182
⚠ This notebook has stopped taking notes: newest record is 23 days old (eng-0280), against an expected cadence of 14 days.

cfg-0182

citable URL: https://halobench.com/records/cfg-0182/ — this address never moves; the anchor /records/#cfg-0182 keeps resolving
modelQwen3.8-Flash-Next-ROCmFP4-FAST-v2-ple16.gguf · ROCmFP4-FAST-v2-ple16 · rev 552a7a162f6a620c3aa0850d070086bc2b95094e0a4e8b860694c7f212cb59d8
engineLaurentZuijdwijk/llama.cpp 5e085d123eead2e89b5c19f824fccb05727da6a2 · vulkan · host aihydra (igpu)
flagsllama-server -m Qwen3.8-Flash-Next-ROCmFP4-FAST-v2-ple16.gguf -ngl 99 -fa on -c 200192 -b 2048 -ub 2048 -ctk q8_0 -ctv q8_0 --parallel 1 --jinja --metrics --spec-type draft-mtp -md Qwen3.8-Flash-Next-MTP-ROCmFP4-FAST.gguf --spec-draft-adaptive; env GGML_VK_ALLOW_GRAPHICS_QUEUE=1; port 5830. The ~51B-parameter n-gram/PLE table (per_layer_token_embd) is baked per-head into the ple16 quant and stays fully GPU-resident (no -ot offload needed). Capability run (run-0653) used -c 196608 directly on llama-server; the decode-depth ladder (run-0654..run-0646) ran through the slotpin energy-metering proxy at -c 200192 (near-identical). tau2 agent args temperature=0.0, max_tokens=4096, airline seed 42, tasks 0-25, claude-haiku-4.5 simulator (sim_self_play=false), max-steps 200, single slot (concurrency 1) with adaptive MTP.
templatenot recorded at test time
treeupstream — stock

Capability basis

measured on this config

production node(s): qwen38-flash-next

Runs on this config (5)

rundatekindsuitemetrics
run-06532026-09-02capabilitytau2-airline-full26tasks_passed=21 · tasks_total=25
run-06542026-09-02performancedecode-depth-serveddecode_tps=16.1 · prefill_tps=298
run-06552026-09-02performancedecode-depth-serveddecode_tps=16 · prefill_tps=219
run-06562026-09-02performancedecode-depth-serveddecode_tps=22.8 · prefill_tps=169
run-06572026-09-02performancedecode-depth-serveddecode_tps=26.1 · prefill_tps=127

Cited by — computed at build time, never stored

claims clm-0128
model pages qwen38-flash-next