Home › Evidence › Records › clm-0077

clm-0077

measured-herehigh ●●●
citable URL: https://halobench.com/records/clm-0077/ — this address never moves; the anchor /records/#clm-0077 keeps resolving

Deep-Thought-Posttrain's KV cache costs a measured 40.00 KiB/token (f16 KV, ROCm and Vulkan identical — both default to the same type_k/type_v) — reproduced exactly from first principles against the GGUF's own attention metadata (2 x 32 layers x 5 KV heads x 64 head_dim x 2 bytes) and cross-checked against a real two-point GTT probe (c=2048 vs c=8192, /sys/class/drm/card0/device/mem_info_gtt_used readings before and after each load): (1,258,139,648 - 1,006,481,408) bytes / (8192-2048) tokens = 40,960 bytes/token exactly. At this candidate's full native context (8192 tokens, the ceiling — not a chosen serving depth), total KV footprint is 0.3125 GiB against a 0.6740 GiB weights footprint on a 120 GiB pool: this candidate poses no realistic fit or OOM risk on this hardware at any context it can actually use, which is a genuinely uninteresting finding in isolation, but sets the honest baseline the site's larger candidates are compared against.

verified 2026-08-17 · volatility low

Note — the record's own working

Two independent instruments agree exactly (empirical sysfs delta vs first-principles GGUF-metadata arithmetic), which is the highest-confidence pattern this lab's KV-cost claims can reach — same standard as ornith-35b's clm-0071.

Cited by — computed at build time, never stored

model pages deep-thought-posttrain
candidate gate history deep-thought-posttrain