citable URL: https://halobench.com/records/clm-0071/ — this address never moves; the anchor /records/#clm-0071 keeps resolving
Ornith-1.0-35B's KV cache costs 20.0 KiB per token on llama.cpp at f16 (default KV, ROCm), measured from a two-point GTT load probe (c=4096 -> 36,750,315,520 bytes gtt_used, c=131,072 -> 39,350,784,000 bytes gtt_used; delta 2,600,468,480 bytes over 126,976 tokens). This is SMALLER than laguna-s-21's 48.0 KiB/token, nemotron3-super's 8.00 KiB/token is still smaller, and deepseek-v4-flash's 7.13 KiB/token smaller still -- consistent with this candidate's hybrid architecture, where only 10 of 40 blocks (full_attention_interval=4) carry standard GQA attention (head_count_kv=2, key/value length 256) and the other 30 hold SSM/linear-attention state that does not grow with context. At the matrix's deepest tested depth (c=131,072), projected GTT use is ~36.6 GiB against the 120 GiB (122,880 MiB) boot window -- comfortably inside, with ~83 GiB headroom, and nowhere near the fit ceiling.
verified 2026-08-16 · volatility low
Note — the record's own working
METHOD: two-point GTT probe on stock 3653e6d ROCm, Ornith-1.0-35B Q8_0, --load-mode none, default f16 KV, gtt_used sampled from /sys/class/drm/card0/device/mem_info_gtt_used immediately after the server reported healthy, server torn down between points. RAW (kvprobe-raw.tsv, aihydra ~/bench-results/ornith-35b-fullbench/): c=4096 -> gtt_used_after=36,947,447,808 bytes; c=131,072 -> gtt_used_after=39,547,916,288 bytes; delta 2,600,468,480 bytes = 2,539,520 KiB / 126,976 tok = 20.00 KiB/token exactly.
INDEPENDENTLY REPRODUCED same session, second probe run with a settling convention (kvprobe-c4096-redo.log / kvprobe-c131072-redo.log): c=4096 -> gtt_used_after_settle=36,750,315,520 bytes; c=131,072 -> gtt_used_after_settle=39,350,784,000 bytes. Absolute readings differ from the first probe by ~197 MiB at BOTH points (idle-baseline drift between the two probe runs, ~3 minutes apart), but the DELTA is bit-for-bit identical: 2,600,468,480 bytes in both runs. Confidence raised to high on this exact reproduction across two independent fresh-process probe pairs -- stronger evidentiary standard than clm-0060/clm-0066's single-pair method, since the confound (idle-baseline drift) is visible in the raw readings and demonstrably does not touch the delta.
CROSS-CHECK: fixed (weights + graph) component at c=4096 = 36,947,447,808 bytes - (4,096 tok * 20.00 KiB/tok * 1024) = 36,947,447,808 - 83,886,080 = 36,863,561,728 bytes = 34.331 GiB, against the GGUF's own model_size (36,903,138,880 bytes = 34.371 GiB per the candidate record's acquisition sha256 verification) -- agrees to within 0.040 GiB (compute-graph + SSM-state overhead), tighter agreement than clm-0060's laguna cross-check.
ARCHITECTURE BASIS for the small per-token cost: this job's own from-scratch GGUF header parser found qwen35moe.full_attention_interval=4 (10 of 40 blocks are full-attention: block ids 3,7,11,15,19,23,27,31,35,39; the other 30 carry ssm_a/ssm_alpha/ssm_beta/ssm_conv1d/ssm_dt.bias/ssm_norm/ssm_out tensors instead of attn_q/k/v/output). KV per token per full-attention layer at f16 = head_count_kv(2) * (key_length(256) + value_length(256)) * 2 bytes = 2048 bytes; x10 layers = 20,480 bytes/token = 20.00 KiB/token -- matches the measured delta exactly, a full first-principles reconciliation rather than just an order-of-magnitude check.