{
  "ranking": [
    {
      "config_id": "984f277c83dc2ee9e8",
      "timestamp": "2026-10-06T02:49:42Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen38-27b-q4km",
      "model_name": "Qwen3.8-27B (identity checked against GGUF)",
      "model_quant": "Q4_K_M",
      "runtime": "ik",
      "phase": "threads",
      "threads": "40",
      "threads_reported": "40",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "",
      "hf_revision": "",
      "placement": "auto",
      "load_mode": "auto",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "7.031088",
      "stddev_tok_s": "0.7658",
      "repetitions": "3",
      "model_bytes": "16102434816",
      "model_params": "26895998464",
      "peak_rss_gib": "15.441158294677734",
      "wall_s": "74.03313839600014",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "/opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/qwen38-27b-q4km.gguf -t 40 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 1 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "f85e2a3fc16e5d3242",
      "timestamp": "2026-10-06T18:22:48Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "gptoss-120b-mxfp4",
      "model_name": "gpt-oss-120b",
      "model_quant": "MXFP4",
      "runtime": "ik",
      "phase": "final-matched-7.0.14-20-pve",
      "threads": "16",
      "threads_reported": "16",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "ggml-org/gpt-oss-120b-GGUF",
      "hf_revision": "238abdd290bb874b90a5da1b4549881b7d05c091",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "on",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "19.971014",
      "stddev_tok_s": "0.046675",
      "repetitions": "3",
      "model_bytes": "63374323968",
      "model_params": "116829156672",
      "peak_rss_gib": "59.259063720703125",
      "wall_s": "87.16584686900023",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35 --membind=1 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/gptoss-120b-mxfp4/gpt-oss-120b-MXFP4.gguf -t 16 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1 -fa 1"
    },
    {
      "config_id": "bce6958046a8d3d409",
      "timestamp": "2026-10-05T19:58:10Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "ik",
      "phase": "numa-controlled",
      "threads": "16",
      "threads_reported": "16",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "26.469284",
      "stddev_tok_s": "0.053596",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.93153381347656",
      "wall_s": "29.54458321799757",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35 --membind=1 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 16 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "3c0e398d058806557e",
      "timestamp": "2026-10-07T02:12:59Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen36-35b-q4kxl",
      "model_name": "Qwen3.6-35B-A3B",
      "model_quant": "UD-Q4_K_XL",
      "runtime": "ik",
      "phase": "pinned-threads",
      "threads": "20",
      "threads_reported": "20",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
      "hf_revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
      "placement": "physical",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "26.563479",
      "stddev_tok_s": "0.117014",
      "repetitions": "3",
      "model_bytes": "22349466112",
      "model_params": "34660610688",
      "peak_rss_gib": "21.046875",
      "wall_s": "25.163988137006527",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=0-39 --interleave=all /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/qwen36-35b-a3b/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -t 20 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "472a78c8ca37eb4ae6",
      "timestamp": "2026-10-06T17:13:37Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen35-35b-q4ks",
      "model_name": "Qwen3.5-35B-A3B",
      "model_quant": "Q4_K_S",
      "runtime": "ik",
      "phase": "final-confirm-7.0.14-20-pve",
      "threads": "20",
      "threads_reported": "20",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3.5-35B-A3B-GGUF",
      "hf_revision": "bc014a17be43adabd7066b7a86075ff935c6a4e2",
      "placement": "physical",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "28.644531",
      "stddev_tok_s": "0.023527",
      "repetitions": "3",
      "model_bytes": "20662856192",
      "model_params": "34660610688",
      "peak_rss_gib": "19.473785400390625",
      "wall_s": "23.72272724600043",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=0-39 --interleave=all /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen35-35b-q4ks/Qwen3.5-35B-A3B-Q4_K_S.gguf -t 20 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "0ecb400e61ed9dbea5",
      "timestamp": "2026-10-05T17:35:34Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "gptoss-20b-q4km",
      "model_name": "gpt-oss-20b",
      "model_quant": "Q4_K_M",
      "runtime": "ik",
      "phase": "pinned-threads",
      "threads": "20",
      "threads_reported": "20",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/gpt-oss-20b-GGUF",
      "hf_revision": "d449b42d93e1c2c7bda5312f5c25c8fb91dfa9b4",
      "placement": "physical",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "33.05084",
      "stddev_tok_s": "0.01315",
      "repetitions": "3",
      "model_bytes": "11611750656",
      "model_params": "20914757184",
      "peak_rss_gib": "11.03757095336914",
      "wall_s": "18.995088342999225",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=0-39 --interleave=all /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/gptoss-20b-q4km/gpt-oss-20b-Q4_K_M.gguf -t 20 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "70675e19e778a871ca",
      "timestamp": "2026-10-06T15:16:36Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3-30b-q4km",
      "model_name": "Qwen3-30B-A3B",
      "model_quant": "Q4_K_M",
      "runtime": "ik",
      "phase": "numa-controlled",
      "threads": "20",
      "threads_reported": "20",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-30B-A3B-GGUF",
      "hf_revision": "d5b1d57bd0b504ac62ae6c725904e96ef228dc74",
      "placement": "physical",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "34.296698",
      "stddev_tok_s": "0.010398",
      "repetitions": "3",
      "model_bytes": "18550716416",
      "model_params": "30532122624",
      "peak_rss_gib": "17.46807861328125",
      "wall_s": "20.82037265699182",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=0-39 --interleave=all /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3-30b-q4km/Qwen3-30B-A3B-Q4_K_M.gguf -t 20 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    }
  ],
  "runtime": [
    {
      "config_id": "f85e2a3fc16e5d3242",
      "timestamp": "2026-10-06T18:22:48Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "gptoss-120b-mxfp4",
      "model_name": "gpt-oss-120b",
      "model_quant": "MXFP4",
      "runtime": "ik",
      "phase": "final-matched-7.0.14-20-pve",
      "threads": "16",
      "threads_reported": "16",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "ggml-org/gpt-oss-120b-GGUF",
      "hf_revision": "238abdd290bb874b90a5da1b4549881b7d05c091",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "on",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "19.971014",
      "stddev_tok_s": "0.046675",
      "repetitions": "3",
      "model_bytes": "63374323968",
      "model_params": "116829156672",
      "peak_rss_gib": "59.259063720703125",
      "wall_s": "87.16584686900023",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35 --membind=1 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/gptoss-120b-mxfp4/gpt-oss-120b-MXFP4.gguf -t 16 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1 -fa 1"
    },
    {
      "config_id": "f85e2a3fc16e5d3242",
      "timestamp": "2026-10-06T18:22:41Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "gptoss-120b-mxfp4",
      "model_name": "gpt-oss-120b",
      "model_quant": "MXFP4",
      "runtime": "ik",
      "phase": "final-matched-7.0.14-20-pve",
      "threads": "16",
      "threads_reported": "16",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "ggml-org/gpt-oss-120b-GGUF",
      "hf_revision": "238abdd290bb874b90a5da1b4549881b7d05c091",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "on",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "512",
      "gen_tokens": "0",
      "test_type": "prefill",
      "mean_tok_s": "196.473289",
      "stddev_tok_s": "3.317347",
      "repetitions": "3",
      "model_bytes": "63374323968",
      "model_params": "116829156672",
      "peak_rss_gib": "59.259063720703125",
      "wall_s": "87.16584686900023",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35 --membind=1 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/gptoss-120b-mxfp4/gpt-oss-120b-MXFP4.gguf -t 16 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1 -fa 1"
    },
    {
      "config_id": "ccfc79d0d1d411cfd8",
      "timestamp": "2026-10-06T18:14:27Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "gptoss-120b-mxfp4",
      "model_name": "gpt-oss-120b",
      "model_quant": "MXFP4",
      "runtime": "upstream",
      "phase": "final-matched-7.0.14-20-pve",
      "threads": "16",
      "threads_reported": "16",
      "build_commit": "0faee50",
      "backends": "CPU",
      "hf_repo": "ggml-org/gpt-oss-120b-GGUF",
      "hf_revision": "238abdd290bb874b90a5da1b4549881b7d05c091",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "on",
      "fa_actual": "1",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "16.110216",
      "stddev_tok_s": "0.015172",
      "repetitions": "3",
      "model_bytes": "63374323968",
      "model_params": "116829156672",
      "peak_rss_gib": "59.24005889892578",
      "wall_s": "204.58852113499597",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35 --membind=1 /opt/llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/gptoss-120b-mxfp4/gpt-oss-120b-MXFP4.gguf -t 16 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -lm none -ctk f16 -ctv f16 -d 0 --repack 1 --poll 50 -fa on"
    },
    {
      "config_id": "ccfc79d0d1d411cfd8",
      "timestamp": "2026-10-06T18:13:52Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "gptoss-120b-mxfp4",
      "model_name": "gpt-oss-120b",
      "model_quant": "MXFP4",
      "runtime": "upstream",
      "phase": "final-matched-7.0.14-20-pve",
      "threads": "16",
      "threads_reported": "16",
      "build_commit": "0faee50",
      "backends": "CPU",
      "hf_repo": "ggml-org/gpt-oss-120b-GGUF",
      "hf_revision": "238abdd290bb874b90a5da1b4549881b7d05c091",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "on",
      "fa_actual": "1",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "512",
      "gen_tokens": "0",
      "test_type": "prefill",
      "mean_tok_s": "59.303903",
      "stddev_tok_s": "0.278047",
      "repetitions": "3",
      "model_bytes": "63374323968",
      "model_params": "116829156672",
      "peak_rss_gib": "59.24005889892578",
      "wall_s": "204.58852113499597",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35 --membind=1 /opt/llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/gptoss-120b-mxfp4/gpt-oss-120b-MXFP4.gguf -t 16 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -lm none -ctk f16 -ctv f16 -d 0 --repack 1 --poll 50 -fa on"
    },
    {
      "config_id": "c6a98f13dd2cb2b014",
      "timestamp": "2026-10-06T17:14:02Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen35-35b-q4ks",
      "model_name": "Qwen3.5-35B-A3B",
      "model_quant": "Q4_K_S",
      "runtime": "ik",
      "phase": "final-matched-7.0.14-20-pve",
      "threads": "16",
      "threads_reported": "16",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3.5-35B-A3B-GGUF",
      "hf_revision": "bc014a17be43adabd7066b7a86075ff935c6a4e2",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "on",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "25.388252",
      "stddev_tok_s": "0.008141",
      "repetitions": "3",
      "model_bytes": "20662856192",
      "model_params": "34660610688",
      "peak_rss_gib": "19.481250762939453",
      "wall_s": "25.28255635900132",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35 --membind=1 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen35-35b-q4ks/Qwen3.5-35B-A3B-Q4_K_S.gguf -t 16 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1 -fa 1"
    },
    {
      "config_id": "c6a98f13dd2cb2b014",
      "timestamp": "2026-10-06T17:13:56Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen35-35b-q4ks",
      "model_name": "Qwen3.5-35B-A3B",
      "model_quant": "Q4_K_S",
      "runtime": "ik",
      "phase": "final-matched-7.0.14-20-pve",
      "threads": "16",
      "threads_reported": "16",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3.5-35B-A3B-GGUF",
      "hf_revision": "bc014a17be43adabd7066b7a86075ff935c6a4e2",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "on",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "512",
      "gen_tokens": "0",
      "test_type": "prefill",
      "mean_tok_s": "284.394353",
      "stddev_tok_s": "1.903328",
      "repetitions": "3",
      "model_bytes": "20662856192",
      "model_params": "34660610688",
      "peak_rss_gib": "19.481250762939453",
      "wall_s": "25.28255635900132",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35 --membind=1 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen35-35b-q4ks/Qwen3.5-35B-A3B-Q4_K_S.gguf -t 16 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1 -fa 1"
    },
    {
      "config_id": "71a2fbc19f15e5b104",
      "timestamp": "2026-10-06T17:11:10Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen35-35b-q4ks",
      "model_name": "Qwen3.5-35B-A3B",
      "model_quant": "Q4_K_S",
      "runtime": "upstream",
      "phase": "final-matched-7.0.14-20-pve",
      "threads": "16",
      "threads_reported": "16",
      "build_commit": "0faee50",
      "backends": "CPU",
      "hf_repo": "unsloth/Qwen3.5-35B-A3B-GGUF",
      "hf_revision": "bc014a17be43adabd7066b7a86075ff935c6a4e2",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "on",
      "fa_actual": "1",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "13.001962",
      "stddev_tok_s": "0.007835",
      "repetitions": "3",
      "model_bytes": "20662856192",
      "model_params": "34660610688",
      "peak_rss_gib": "19.48741912841797",
      "wall_s": "83.6674937929929",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35 --membind=1 /opt/llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen35-35b-q4ks/Qwen3.5-35B-A3B-Q4_K_S.gguf -t 16 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -lm none -ctk f16 -ctv f16 -d 0 --repack 1 --poll 50 -fa on"
    },
    {
      "config_id": "71a2fbc19f15e5b104",
      "timestamp": "2026-10-06T17:10:51Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen35-35b-q4ks",
      "model_name": "Qwen3.5-35B-A3B",
      "model_quant": "Q4_K_S",
      "runtime": "upstream",
      "phase": "final-matched-7.0.14-20-pve",
      "threads": "16",
      "threads_reported": "16",
      "build_commit": "0faee50",
      "backends": "CPU",
      "hf_repo": "unsloth/Qwen3.5-35B-A3B-GGUF",
      "hf_revision": "bc014a17be43adabd7066b7a86075ff935c6a4e2",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "on",
      "fa_actual": "1",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "512",
      "gen_tokens": "0",
      "test_type": "prefill",
      "mean_tok_s": "107.274017",
      "stddev_tok_s": "0.35742",
      "repetitions": "3",
      "model_bytes": "20662856192",
      "model_params": "34660610688",
      "peak_rss_gib": "19.48741912841797",
      "wall_s": "83.6674937929929",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35 --membind=1 /opt/llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen35-35b-q4ks/Qwen3.5-35B-A3B-Q4_K_S.gguf -t 16 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -lm none -ctk f16 -ctv f16 -d 0 --repack 1 --poll 50 -fa on"
    },
    {
      "config_id": "9d81c9b3c2baf30747",
      "timestamp": "2026-10-06T17:49:03Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "ik",
      "phase": "final-matched-7.0.14-20-pve",
      "threads": "16",
      "threads_reported": "16",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "on",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "26.02927",
      "stddev_tok_s": "0.059795",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.932613372802734",
      "wall_s": "64.95550032799656",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35 --membind=1 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 16 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1 -fa 1"
    },
    {
      "config_id": "9d81c9b3c2baf30747",
      "timestamp": "2026-10-06T17:48:56Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "ik",
      "phase": "final-matched-7.0.14-20-pve",
      "threads": "16",
      "threads_reported": "16",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "on",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "512",
      "gen_tokens": "0",
      "test_type": "prefill",
      "mean_tok_s": "247.699526",
      "stddev_tok_s": "3.462027",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.932613372802734",
      "wall_s": "64.95550032799656",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35 --membind=1 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 16 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1 -fa 1"
    },
    {
      "config_id": "be09e70fc78ec3a8db",
      "timestamp": "2026-10-06T17:44:06Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "upstream",
      "phase": "final-matched-7.0.14-20-pve",
      "threads": "16",
      "threads_reported": "16",
      "build_commit": "0faee50",
      "backends": "CPU",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "on",
      "fa_actual": "1",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "12.078324",
      "stddev_tok_s": "0.023297",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.96331787109375",
      "wall_s": "116.05930830299621",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35 --membind=1 /opt/llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 16 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -lm none -ctk f16 -ctv f16 -d 0 --repack 1 --poll 50 -fa on"
    },
    {
      "config_id": "be09e70fc78ec3a8db",
      "timestamp": "2026-10-06T17:43:47Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "upstream",
      "phase": "final-matched-7.0.14-20-pve",
      "threads": "16",
      "threads_reported": "16",
      "build_commit": "0faee50",
      "backends": "CPU",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "on",
      "fa_actual": "1",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "512",
      "gen_tokens": "0",
      "test_type": "prefill",
      "mean_tok_s": "109.651355",
      "stddev_tok_s": "1.123216",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.96331787109375",
      "wall_s": "116.05930830299621",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35 --membind=1 /opt/llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 16 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -lm none -ctk f16 -ctv f16 -d 0 --repack 1 --poll 50 -fa on"
    }
  ],
  "threads": [
    {
      "config_id": "1ac7ff99a334bfa0ba",
      "timestamp": "2026-10-05T20:00:49Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "ik",
      "phase": "pinned-threads",
      "threads": "4",
      "threads_reported": "4",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "14.329311",
      "stddev_tok_s": "0.014553",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.932315826416016",
      "wall_s": "58.69519148100517",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23 --membind=1 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 4 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "6049604413310327a4",
      "timestamp": "2026-10-05T20:01:41Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "ik",
      "phase": "pinned-threads",
      "threads": "6",
      "threads_reported": "6",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "18.428597",
      "stddev_tok_s": "0.003366",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.93397521972656",
      "wall_s": "44.9956404800032",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25 --membind=1 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 6 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "2b02c5796bf3ba4b57",
      "timestamp": "2026-10-05T20:02:21Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "ik",
      "phase": "pinned-threads",
      "threads": "8",
      "threads_reported": "8",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "21.736471",
      "stddev_tok_s": "0.105925",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.9333381652832",
      "wall_s": "37.17035391999525",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27 --membind=1 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 8 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "297f36b6f8556af58a",
      "timestamp": "2026-10-05T20:02:58Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "ik",
      "phase": "pinned-threads",
      "threads": "9",
      "threads_reported": "9",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "22.689318",
      "stddev_tok_s": "0.020088",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.933074951171875",
      "wall_s": "36.011769939999795",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28 --membind=1 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 9 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "46afda0027a77bbd3d",
      "timestamp": "2026-10-05T20:03:34Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "ik",
      "phase": "pinned-threads",
      "threads": "10",
      "threads_reported": "10",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "24.13623",
      "stddev_tok_s": "0.027223",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.9334831237793",
      "wall_s": "33.953696911004954",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28,29 --membind=1 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 10 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "24572f86cce405bca5",
      "timestamp": "2026-10-05T20:04:07Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "ik",
      "phase": "pinned-threads",
      "threads": "12",
      "threads_reported": "12",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "25.252223",
      "stddev_tok_s": "0.026932",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.934913635253906",
      "wall_s": "31.74987821999821",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28,29,30,31 --membind=1 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 12 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "41d1d676326f6358fd",
      "timestamp": "2026-10-05T20:04:37Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "ik",
      "phase": "pinned-threads",
      "threads": "16",
      "threads_reported": "16",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "25.936091",
      "stddev_tok_s": "0.02702",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.93208694458008",
      "wall_s": "29.695454884000355",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35 --membind=1 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 16 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "afe8502f2b6f498055",
      "timestamp": "2026-10-05T20:05:06Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "ik",
      "phase": "pinned-threads",
      "threads": "20",
      "threads_reported": "20",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "25.086178",
      "stddev_tok_s": "0.031544",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.93262481689453",
      "wall_s": "29.151571260998026",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39 --membind=1 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 20 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "8ff6e181ed783761c5",
      "timestamp": "2026-10-05T19:47:20Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "ik",
      "phase": "threads",
      "threads": "4",
      "threads_reported": "4",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "auto",
      "load_mode": "auto",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "11.640066",
      "stddev_tok_s": "0.600741",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.91375732421875",
      "wall_s": "59.19622165799956",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "/opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 4 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 1 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "18ba739132f95bcb89",
      "timestamp": "2026-10-05T19:48:52Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "ik",
      "phase": "threads",
      "threads": "6",
      "threads_reported": "6",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "auto",
      "load_mode": "auto",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "14.125792",
      "stddev_tok_s": "0.384761",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.91764831542969",
      "wall_s": "48.00790587799929",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "/opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 6 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 1 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "975fe9acae8d0fad51",
      "timestamp": "2026-10-05T19:45:55Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "ik",
      "phase": "threads",
      "threads": "8",
      "threads_reported": "8",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "auto",
      "load_mode": "auto",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "16.348417",
      "stddev_tok_s": "0.141701",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.91559982299805",
      "wall_s": "42.28306683700066",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "/opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 8 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 1 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "1d331d1bda2223f076",
      "timestamp": "2026-10-05T19:50:13Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "ik",
      "phase": "threads",
      "threads": "9",
      "threads_reported": "9",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "auto",
      "load_mode": "auto",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "17.037708",
      "stddev_tok_s": "0.172376",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.91820526123047",
      "wall_s": "40.8290306639974",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "/opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 9 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 1 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "7aa242c364dbed69ee",
      "timestamp": "2026-10-05T19:50:54Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "ik",
      "phase": "threads",
      "threads": "10",
      "threads_reported": "10",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "auto",
      "load_mode": "auto",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "17.495074",
      "stddev_tok_s": "0.097883",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.9162483215332",
      "wall_s": "39.376529844998004",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "/opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 10 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 1 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "e6452b467904f2b6e4",
      "timestamp": "2026-10-05T19:48:10Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "ik",
      "phase": "threads",
      "threads": "12",
      "threads_reported": "12",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "auto",
      "load_mode": "auto",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "18.983964",
      "stddev_tok_s": "0.362996",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.91816329956055",
      "wall_s": "36.864416775999416",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "/opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 12 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 1 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "7ac59ffea1136400d0",
      "timestamp": "2026-10-05T19:46:34Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "ik",
      "phase": "threads",
      "threads": "16",
      "threads_reported": "16",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "auto",
      "load_mode": "auto",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "19.78435",
      "stddev_tok_s": "0.278891",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.9151496887207",
      "wall_s": "34.66374312199696",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "/opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 16 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 1 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "537c76a9a18da5e555",
      "timestamp": "2026-10-05T19:49:34Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "ik",
      "phase": "threads",
      "threads": "20",
      "threads_reported": "20",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "auto",
      "load_mode": "auto",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "19.067055",
      "stddev_tok_s": "0.958661",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.91852951049805",
      "wall_s": "35.318710068000655",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "/opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 20 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 1 -ctk f16 -ctv f16 -rtr 1"
    }
  ],
  "numa": {
    "8": [
      {
        "config_id": "ec99a05d7144bd3a7c",
        "timestamp": "2026-10-07T02:07:14Z",
        "kernel_release": "7.0.14-20-pve",
        "model_id": "qwen36-35b-q4kxl",
        "model_name": "Qwen3.6-35B-A3B",
        "model_quant": "UD-Q4_K_XL",
        "runtime": "ik",
        "phase": "numa-controlled",
        "threads": "8",
        "threads_reported": "8",
        "build_commit": "unknown",
        "backends": "",
        "hf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
        "hf_revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
        "placement": "cross",
        "load_mode": "none",
        "batch_size": "2048",
        "ubatch_size": "512",
        "kv_k": "f16",
        "kv_v": "f16",
        "fa_requested": "auto",
        "fa_actual": "True",
        "repack_requested": "1",
        "repack_actual": "True",
        "depth_tokens": "0",
        "prompt_tokens": "0",
        "gen_tokens": "128",
        "test_type": "decode",
        "mean_tok_s": "18.641095",
        "stddev_tok_s": "0.092783",
        "repetitions": "3",
        "model_bytes": "22349466112",
        "model_params": "34660610688",
        "peak_rss_gib": "21.05292510986328",
        "wall_s": "36.05574414899456",
        "protocol_version": "2",
        "status": "ok",
        "error": "",
        "command": "numactl --physcpubind=0,1,2,3,20,21,22,23 --interleave=all /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/qwen36-35b-a3b/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -t 8 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
      },
      {
        "config_id": "96ef0bd499109de3fa",
        "timestamp": "2026-10-07T02:04:10Z",
        "kernel_release": "7.0.14-20-pve",
        "model_id": "qwen36-35b-q4kxl",
        "model_name": "Qwen3.6-35B-A3B",
        "model_quant": "UD-Q4_K_XL",
        "runtime": "ik",
        "phase": "numa-controlled",
        "threads": "8",
        "threads_reported": "8",
        "build_commit": "unknown",
        "backends": "",
        "hf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
        "hf_revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
        "placement": "node0",
        "load_mode": "none",
        "batch_size": "2048",
        "ubatch_size": "512",
        "kv_k": "f16",
        "kv_v": "f16",
        "fa_requested": "auto",
        "fa_actual": "True",
        "repack_requested": "1",
        "repack_actual": "True",
        "depth_tokens": "0",
        "prompt_tokens": "0",
        "gen_tokens": "128",
        "test_type": "decode",
        "mean_tok_s": "19.874309",
        "stddev_tok_s": "0.151361",
        "repetitions": "3",
        "model_bytes": "22349466112",
        "model_params": "34660610688",
        "peak_rss_gib": "21.051918029785156",
        "wall_s": "34.31643210200127",
        "protocol_version": "2",
        "status": "ok",
        "error": "",
        "command": "numactl --physcpubind=0,1,2,3,4,5,6,7 --membind=0 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/qwen36-35b-a3b/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -t 8 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
      },
      {
        "config_id": "e5bc2c60ff06cb4918",
        "timestamp": "2026-10-07T02:00:57Z",
        "kernel_release": "7.0.14-20-pve",
        "model_id": "qwen36-35b-q4kxl",
        "model_name": "Qwen3.6-35B-A3B",
        "model_quant": "UD-Q4_K_XL",
        "runtime": "ik",
        "phase": "numa-controlled",
        "threads": "8",
        "threads_reported": "8",
        "build_commit": "unknown",
        "backends": "",
        "hf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
        "hf_revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
        "placement": "auto",
        "load_mode": "none",
        "batch_size": "2048",
        "ubatch_size": "512",
        "kv_k": "f16",
        "kv_v": "f16",
        "fa_requested": "auto",
        "fa_actual": "True",
        "repack_requested": "1",
        "repack_actual": "True",
        "depth_tokens": "0",
        "prompt_tokens": "0",
        "gen_tokens": "128",
        "test_type": "decode",
        "mean_tok_s": "16.125603",
        "stddev_tok_s": "0.420239",
        "repetitions": "3",
        "model_bytes": "22349466112",
        "model_params": "34660610688",
        "peak_rss_gib": "21.03414535522461",
        "wall_s": "38.009356280002976",
        "protocol_version": "2",
        "status": "ok",
        "error": "",
        "command": "/opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/qwen36-35b-a3b/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -t 8 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
      },
      {
        "config_id": "118bdf5e97e918b6c3",
        "timestamp": "2026-10-07T02:05:43Z",
        "kernel_release": "7.0.14-20-pve",
        "model_id": "qwen36-35b-q4kxl",
        "model_name": "Qwen3.6-35B-A3B",
        "model_quant": "UD-Q4_K_XL",
        "runtime": "ik",
        "phase": "numa-controlled",
        "threads": "8",
        "threads_reported": "8",
        "build_commit": "unknown",
        "backends": "",
        "hf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
        "hf_revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
        "placement": "node1",
        "load_mode": "none",
        "batch_size": "2048",
        "ubatch_size": "512",
        "kv_k": "f16",
        "kv_v": "f16",
        "fa_requested": "auto",
        "fa_actual": "True",
        "repack_requested": "1",
        "repack_actual": "True",
        "depth_tokens": "0",
        "prompt_tokens": "0",
        "gen_tokens": "128",
        "test_type": "decode",
        "mean_tok_s": "21.119928",
        "stddev_tok_s": "0.01198",
        "repetitions": "3",
        "model_bytes": "22349466112",
        "model_params": "34660610688",
        "peak_rss_gib": "21.052814483642578",
        "wall_s": "32.07508965098532",
        "protocol_version": "2",
        "status": "ok",
        "error": "",
        "command": "numactl --physcpubind=20,21,22,23,24,25,26,27 --membind=1 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/qwen36-35b-a3b/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -t 8 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
      },
      {
        "config_id": "a223d25cf3ba01e1af",
        "timestamp": "2026-10-07T02:02:38Z",
        "kernel_release": "7.0.14-20-pve",
        "model_id": "qwen36-35b-q4kxl",
        "model_name": "Qwen3.6-35B-A3B",
        "model_quant": "UD-Q4_K_XL",
        "runtime": "ik",
        "phase": "numa-controlled",
        "threads": "8",
        "threads_reported": "8",
        "build_commit": "unknown",
        "backends": "",
        "hf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
        "hf_revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
        "placement": "physical",
        "load_mode": "none",
        "batch_size": "2048",
        "ubatch_size": "512",
        "kv_k": "f16",
        "kv_v": "f16",
        "fa_requested": "auto",
        "fa_actual": "True",
        "repack_requested": "1",
        "repack_actual": "True",
        "depth_tokens": "0",
        "prompt_tokens": "0",
        "gen_tokens": "128",
        "test_type": "decode",
        "mean_tok_s": "18.430726",
        "stddev_tok_s": "0.014875",
        "repetitions": "3",
        "model_bytes": "22349466112",
        "model_params": "34660610688",
        "peak_rss_gib": "21.045520782470703",
        "wall_s": "35.294171115994686",
        "protocol_version": "2",
        "status": "ok",
        "error": "",
        "command": "numactl --physcpubind=0-39 --interleave=all /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/qwen36-35b-a3b/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -t 8 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
      }
    ],
    "12": [
      {
        "config_id": "490cb887a688c05481",
        "timestamp": "2026-10-07T02:01:34Z",
        "kernel_release": "7.0.14-20-pve",
        "model_id": "qwen36-35b-q4kxl",
        "model_name": "Qwen3.6-35B-A3B",
        "model_quant": "UD-Q4_K_XL",
        "runtime": "ik",
        "phase": "numa-controlled",
        "threads": "12",
        "threads_reported": "12",
        "build_commit": "unknown",
        "backends": "",
        "hf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
        "hf_revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
        "placement": "auto",
        "load_mode": "none",
        "batch_size": "2048",
        "ubatch_size": "512",
        "kv_k": "f16",
        "kv_v": "f16",
        "fa_requested": "auto",
        "fa_actual": "True",
        "repack_requested": "1",
        "repack_actual": "True",
        "depth_tokens": "0",
        "prompt_tokens": "0",
        "gen_tokens": "128",
        "test_type": "decode",
        "mean_tok_s": "18.928877",
        "stddev_tok_s": "0.365791",
        "repetitions": "3",
        "model_bytes": "22349466112",
        "model_params": "34660610688",
        "peak_rss_gib": "21.039840698242188",
        "wall_s": "32.340351955994265",
        "protocol_version": "2",
        "status": "ok",
        "error": "",
        "command": "/opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/qwen36-35b-a3b/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -t 12 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
      },
      {
        "config_id": "402d39851de93393d9",
        "timestamp": "2026-10-07T02:07:48Z",
        "kernel_release": "7.0.14-20-pve",
        "model_id": "qwen36-35b-q4kxl",
        "model_name": "Qwen3.6-35B-A3B",
        "model_quant": "UD-Q4_K_XL",
        "runtime": "ik",
        "phase": "numa-controlled",
        "threads": "12",
        "threads_reported": "12",
        "build_commit": "unknown",
        "backends": "",
        "hf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
        "hf_revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
        "placement": "cross",
        "load_mode": "none",
        "batch_size": "2048",
        "ubatch_size": "512",
        "kv_k": "f16",
        "kv_v": "f16",
        "fa_requested": "auto",
        "fa_actual": "True",
        "repack_requested": "1",
        "repack_actual": "True",
        "depth_tokens": "0",
        "prompt_tokens": "0",
        "gen_tokens": "128",
        "test_type": "decode",
        "mean_tok_s": "22.883358",
        "stddev_tok_s": "0.061806",
        "repetitions": "3",
        "model_bytes": "22349466112",
        "model_params": "34660610688",
        "peak_rss_gib": "21.052478790283203",
        "wall_s": "29.886228540010052",
        "protocol_version": "2",
        "status": "ok",
        "error": "",
        "command": "numactl --physcpubind=0,1,2,3,4,5,20,21,22,23,24,25 --interleave=all /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/qwen36-35b-a3b/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -t 12 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
      },
      {
        "config_id": "b157aa472e8b73929a",
        "timestamp": "2026-10-07T02:03:12Z",
        "kernel_release": "7.0.14-20-pve",
        "model_id": "qwen36-35b-q4kxl",
        "model_name": "Qwen3.6-35B-A3B",
        "model_quant": "UD-Q4_K_XL",
        "runtime": "ik",
        "phase": "numa-controlled",
        "threads": "12",
        "threads_reported": "12",
        "build_commit": "unknown",
        "backends": "",
        "hf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
        "hf_revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
        "placement": "physical",
        "load_mode": "none",
        "batch_size": "2048",
        "ubatch_size": "512",
        "kv_k": "f16",
        "kv_v": "f16",
        "fa_requested": "auto",
        "fa_actual": "True",
        "repack_requested": "1",
        "repack_actual": "True",
        "depth_tokens": "0",
        "prompt_tokens": "0",
        "gen_tokens": "128",
        "test_type": "decode",
        "mean_tok_s": "22.990064",
        "stddev_tok_s": "0.023938",
        "repetitions": "3",
        "model_bytes": "22349466112",
        "model_params": "34660610688",
        "peak_rss_gib": "21.044574737548828",
        "wall_s": "29.081596010975773",
        "protocol_version": "2",
        "status": "ok",
        "error": "",
        "command": "numactl --physcpubind=0-39 --interleave=all /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/qwen36-35b-a3b/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -t 12 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
      },
      {
        "config_id": "bf67669c5e45dea412",
        "timestamp": "2026-10-07T02:04:43Z",
        "kernel_release": "7.0.14-20-pve",
        "model_id": "qwen36-35b-q4kxl",
        "model_name": "Qwen3.6-35B-A3B",
        "model_quant": "UD-Q4_K_XL",
        "runtime": "ik",
        "phase": "numa-controlled",
        "threads": "12",
        "threads_reported": "12",
        "build_commit": "unknown",
        "backends": "",
        "hf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
        "hf_revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
        "placement": "node0",
        "load_mode": "none",
        "batch_size": "2048",
        "ubatch_size": "512",
        "kv_k": "f16",
        "kv_v": "f16",
        "fa_requested": "auto",
        "fa_actual": "True",
        "repack_requested": "1",
        "repack_actual": "True",
        "depth_tokens": "0",
        "prompt_tokens": "0",
        "gen_tokens": "128",
        "test_type": "decode",
        "mean_tok_s": "22.58053",
        "stddev_tok_s": "0.082463",
        "repetitions": "3",
        "model_bytes": "22349466112",
        "model_params": "34660610688",
        "peak_rss_gib": "21.052608489990234",
        "wall_s": "29.436129390000133",
        "protocol_version": "2",
        "status": "ok",
        "error": "",
        "command": "numactl --physcpubind=0,1,2,3,4,5,6,7,8,9,10,11 --membind=0 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/qwen36-35b-a3b/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -t 12 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
      },
      {
        "config_id": "ed060c8dcbfd608922",
        "timestamp": "2026-10-07T02:06:14Z",
        "kernel_release": "7.0.14-20-pve",
        "model_id": "qwen36-35b-q4kxl",
        "model_name": "Qwen3.6-35B-A3B",
        "model_quant": "UD-Q4_K_XL",
        "runtime": "ik",
        "phase": "numa-controlled",
        "threads": "12",
        "threads_reported": "12",
        "build_commit": "unknown",
        "backends": "",
        "hf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
        "hf_revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
        "placement": "node1",
        "load_mode": "none",
        "batch_size": "2048",
        "ubatch_size": "512",
        "kv_k": "f16",
        "kv_v": "f16",
        "fa_requested": "auto",
        "fa_actual": "True",
        "repack_requested": "1",
        "repack_actual": "True",
        "depth_tokens": "0",
        "prompt_tokens": "0",
        "gen_tokens": "128",
        "test_type": "decode",
        "mean_tok_s": "23.226008",
        "stddev_tok_s": "0.14097",
        "repetitions": "3",
        "model_bytes": "22349466112",
        "model_params": "34660610688",
        "peak_rss_gib": "21.051116943359375",
        "wall_s": "28.31261056099902",
        "protocol_version": "2",
        "status": "ok",
        "error": "",
        "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28,29,30,31 --membind=1 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/qwen36-35b-a3b/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -t 12 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
      }
    ]
  },
  "quantization": [
    {
      "config_id": "5126c947a27bdfc13f",
      "timestamp": "2026-10-05T20:40:04Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen35-35b-q4km",
      "model_name": "Qwen3.5-35B-A3B",
      "model_quant": "Q4_K_M",
      "runtime": "ik",
      "phase": "pinned-threads",
      "threads": "20",
      "threads_reported": "20",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "",
      "hf_revision": "",
      "placement": "physical",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "27.597416",
      "stddev_tok_s": "0.028392",
      "repetitions": "3",
      "model_bytes": "22005033472",
      "model_params": "34660610688",
      "peak_rss_gib": "20.72418975830078",
      "wall_s": "42.82272892900073",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=0-39 --interleave=all /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/qwen35-35b-a3b-q4km.gguf -t 20 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "472a78c8ca37eb4ae6",
      "timestamp": "2026-10-06T17:13:37Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen35-35b-q4ks",
      "model_name": "Qwen3.5-35B-A3B",
      "model_quant": "Q4_K_S",
      "runtime": "ik",
      "phase": "final-confirm-7.0.14-20-pve",
      "threads": "20",
      "threads_reported": "20",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3.5-35B-A3B-GGUF",
      "hf_revision": "bc014a17be43adabd7066b7a86075ff935c6a4e2",
      "placement": "physical",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "28.644531",
      "stddev_tok_s": "0.023527",
      "repetitions": "3",
      "model_bytes": "20662856192",
      "model_params": "34660610688",
      "peak_rss_gib": "19.473785400390625",
      "wall_s": "23.72272724600043",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=0-39 --interleave=all /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen35-35b-q4ks/Qwen3.5-35B-A3B-Q4_K_S.gguf -t 20 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "c841d7546283ddca58",
      "timestamp": "2026-10-06T10:02:44Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen35-35b-q5km",
      "model_name": "Qwen3.5-35B-A3B",
      "model_quant": "Q5_K_M",
      "runtime": "ik",
      "phase": "pinned-threads",
      "threads": "20",
      "threads_reported": "20",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3.5-35B-A3B-GGUF",
      "hf_revision": "bc014a17be43adabd7066b7a86075ff935c6a4e2",
      "placement": "physical",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "25.777639",
      "stddev_tok_s": "0.23821",
      "repetitions": "3",
      "model_bytes": "26238618112",
      "model_params": "34660610688",
      "peak_rss_gib": "24.66645050048828",
      "wall_s": "26.597901722008828",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=0-39 --interleave=all /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen35-35b-q5km/Qwen3.5-35B-A3B-Q5_K_M.gguf -t 20 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "5bb2037057d1ac511d",
      "timestamp": "2026-10-06T13:10:22Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen35-35b-q80",
      "model_name": "Qwen3.5-35B-A3B",
      "model_quant": "Q8_0",
      "runtime": "ik",
      "phase": "pinned-threads",
      "threads": "20",
      "threads_reported": "20",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3.5-35B-A3B-GGUF",
      "hf_revision": "bc014a17be43adabd7066b7a86075ff935c6a4e2",
      "placement": "physical",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "22.684624",
      "stddev_tok_s": "0.045165",
      "repetitions": "3",
      "model_bytes": "36892150272",
      "model_params": "34660610688",
      "peak_rss_gib": "34.58561706542969",
      "wall_s": "30.387182698003016",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=0-39 --interleave=all /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen35-35b-q80/Qwen3.5-35B-A3B-Q8_0.gguf -t 20 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "5657664df4c10d474f",
      "timestamp": "2026-10-05T12:06:28Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen35-35b-ud-iq4xs",
      "model_name": "Qwen3.5-35B-A3B",
      "model_quant": "UD-IQ4_XS",
      "runtime": "ik",
      "phase": "numa-controlled",
      "threads": "20",
      "threads_reported": "20",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3.5-35B-A3B-GGUF",
      "hf_revision": "bc014a17be43adabd7066b7a86075ff935c6a4e2",
      "placement": "physical",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "24.462877",
      "stddev_tok_s": "0.094613",
      "repetitions": "3",
      "model_bytes": "17475185152",
      "model_params": "34660610688",
      "peak_rss_gib": "16.50463104248047",
      "wall_s": "38.5785863279998",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=0-39 --interleave=all /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen35-35b-ud-iq4xs/Qwen3.5-35B-A3B-UD-IQ4_XS.gguf -t 20 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    }
  ],
  "frontier": [
    {
      "config_id": "984f277c83dc2ee9e8",
      "timestamp": "2026-10-06T02:49:42Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen38-27b-q4km",
      "model_name": "Qwen3.8-27B (identity checked against GGUF)",
      "model_quant": "Q4_K_M",
      "runtime": "ik",
      "phase": "threads",
      "threads": "40",
      "threads_reported": "40",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "",
      "hf_revision": "",
      "placement": "auto",
      "load_mode": "auto",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "7.031088",
      "stddev_tok_s": "0.7658",
      "repetitions": "3",
      "model_bytes": "16102434816",
      "model_params": "26895998464",
      "peak_rss_gib": "15.441158294677734",
      "wall_s": "74.03313839600014",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "/opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/qwen38-27b-q4km.gguf -t 40 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 1 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "f85e2a3fc16e5d3242",
      "timestamp": "2026-10-06T18:22:48Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "gptoss-120b-mxfp4",
      "model_name": "gpt-oss-120b",
      "model_quant": "MXFP4",
      "runtime": "ik",
      "phase": "final-matched-7.0.14-20-pve",
      "threads": "16",
      "threads_reported": "16",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "ggml-org/gpt-oss-120b-GGUF",
      "hf_revision": "238abdd290bb874b90a5da1b4549881b7d05c091",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "on",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "19.971014",
      "stddev_tok_s": "0.046675",
      "repetitions": "3",
      "model_bytes": "63374323968",
      "model_params": "116829156672",
      "peak_rss_gib": "59.259063720703125",
      "wall_s": "87.16584686900023",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35 --membind=1 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/gptoss-120b-mxfp4/gpt-oss-120b-MXFP4.gguf -t 16 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1 -fa 1"
    },
    {
      "config_id": "bce6958046a8d3d409",
      "timestamp": "2026-10-05T19:58:10Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3next-80b-iq4xs",
      "model_name": "Qwen3-Next-80B-A3B-Instruct",
      "model_quant": "IQ4_XS",
      "runtime": "ik",
      "phase": "numa-controlled",
      "threads": "16",
      "threads_reported": "16",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "hf_revision": "6688872b7a4e5eb8a121ceea08009530a024ba1d",
      "placement": "node1",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "26.469284",
      "stddev_tok_s": "0.053596",
      "repetitions": "3",
      "model_bytes": "42638492672",
      "model_params": "79674391296",
      "peak_rss_gib": "39.93153381347656",
      "wall_s": "29.54458321799757",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35 --membind=1 /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3next-80b-iq4xs/Qwen3-Next-80B-A3B-Instruct-IQ4_XS.gguf -t 16 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "3c0e398d058806557e",
      "timestamp": "2026-10-07T02:12:59Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen36-35b-q4kxl",
      "model_name": "Qwen3.6-35B-A3B",
      "model_quant": "UD-Q4_K_XL",
      "runtime": "ik",
      "phase": "pinned-threads",
      "threads": "20",
      "threads_reported": "20",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
      "hf_revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
      "placement": "physical",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "26.563479",
      "stddev_tok_s": "0.117014",
      "repetitions": "3",
      "model_bytes": "22349466112",
      "model_params": "34660610688",
      "peak_rss_gib": "21.046875",
      "wall_s": "25.163988137006527",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=0-39 --interleave=all /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/qwen36-35b-a3b/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -t 20 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "472a78c8ca37eb4ae6",
      "timestamp": "2026-10-06T17:13:37Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen35-35b-q4ks",
      "model_name": "Qwen3.5-35B-A3B",
      "model_quant": "Q4_K_S",
      "runtime": "ik",
      "phase": "final-confirm-7.0.14-20-pve",
      "threads": "20",
      "threads_reported": "20",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3.5-35B-A3B-GGUF",
      "hf_revision": "bc014a17be43adabd7066b7a86075ff935c6a4e2",
      "placement": "physical",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "28.644531",
      "stddev_tok_s": "0.023527",
      "repetitions": "3",
      "model_bytes": "20662856192",
      "model_params": "34660610688",
      "peak_rss_gib": "19.473785400390625",
      "wall_s": "23.72272724600043",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=0-39 --interleave=all /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen35-35b-q4ks/Qwen3.5-35B-A3B-Q4_K_S.gguf -t 20 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "0ecb400e61ed9dbea5",
      "timestamp": "2026-10-05T17:35:34Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "gptoss-20b-q4km",
      "model_name": "gpt-oss-20b",
      "model_quant": "Q4_K_M",
      "runtime": "ik",
      "phase": "pinned-threads",
      "threads": "20",
      "threads_reported": "20",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/gpt-oss-20b-GGUF",
      "hf_revision": "d449b42d93e1c2c7bda5312f5c25c8fb91dfa9b4",
      "placement": "physical",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "33.05084",
      "stddev_tok_s": "0.01315",
      "repetitions": "3",
      "model_bytes": "11611750656",
      "model_params": "20914757184",
      "peak_rss_gib": "11.03757095336914",
      "wall_s": "18.995088342999225",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=0-39 --interleave=all /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/gptoss-20b-q4km/gpt-oss-20b-Q4_K_M.gguf -t 20 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    },
    {
      "config_id": "70675e19e778a871ca",
      "timestamp": "2026-10-06T15:16:36Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen3-30b-q4km",
      "model_name": "Qwen3-30B-A3B",
      "model_quant": "Q4_K_M",
      "runtime": "ik",
      "phase": "numa-controlled",
      "threads": "20",
      "threads_reported": "20",
      "build_commit": "unknown",
      "backends": "",
      "hf_repo": "unsloth/Qwen3-30B-A3B-GGUF",
      "hf_revision": "d5b1d57bd0b504ac62ae6c725904e96ef228dc74",
      "placement": "physical",
      "load_mode": "none",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "True",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode",
      "mean_tok_s": "34.296698",
      "stddev_tok_s": "0.010398",
      "repetitions": "3",
      "model_bytes": "18550716416",
      "model_params": "30532122624",
      "peak_rss_gib": "17.46807861328125",
      "wall_s": "20.82037265699182",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=0-39 --interleave=all /opt/llm-benchmarks/pve01-20261004/ik_llama.cpp/build/bin/llama-bench -m /models/llm-campaign-20261004/qwen3-30b-q4km/Qwen3-30B-A3B-Q4_K_M.gguf -t 20 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -mmp 0 -ctk f16 -ctv f16 -rtr 1"
    }
  ],
  "context": [
    {
      "config_id": "ffbd90e88bae3b2aae",
      "timestamp": "2026-10-07T00:38:47Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen36-35b-q4kxl",
      "model_name": "Qwen3.6-35B-A3B",
      "model_quant": "UD-Q4_K_XL",
      "runtime": "upstream",
      "phase": "context",
      "threads": "8",
      "threads_reported": "8",
      "build_commit": "0faee50",
      "backends": "CPU",
      "hf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
      "hf_revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
      "placement": "node1",
      "load_mode": "mlock",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "q8_0",
      "kv_v": "q8_0",
      "fa_requested": "on",
      "fa_actual": "1",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "8192",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode_context",
      "mean_tok_s": "9.248489",
      "stddev_tok_s": "0.005863",
      "repetitions": "3",
      "model_bytes": "22349466112",
      "model_params": "34660610688",
      "peak_rss_gib": "21.30107879638672",
      "wall_s": "269.49904504799633",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27 --membind=1 /opt/llama.cpp/build/bin/llama-bench -m /models/qwen36-35b-a3b/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -t 8 -p 0 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -lm mlock -ctk q8_0 -ctv q8_0 -d 8192 --repack 1 --poll 50 -fa on"
    },
    {
      "config_id": "3e59d5c38fd636d400",
      "timestamp": "2026-10-07T00:33:11Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen36-35b-q4kxl",
      "model_name": "Qwen3.6-35B-A3B",
      "model_quant": "UD-Q4_K_XL",
      "runtime": "upstream",
      "phase": "context",
      "threads": "8",
      "threads_reported": "8",
      "build_commit": "0faee50",
      "backends": "CPU",
      "hf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
      "hf_revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
      "placement": "node1",
      "load_mode": "mlock",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "q8_0",
      "kv_v": "q8_0",
      "fa_requested": "on",
      "fa_actual": "1",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "4096",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode_context",
      "mean_tok_s": "10.392604",
      "stddev_tok_s": "0.012723",
      "repetitions": "3",
      "model_bytes": "22349466112",
      "model_params": "34660610688",
      "peak_rss_gib": "21.214706420898438",
      "wall_s": "143.39320216499618",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27 --membind=1 /opt/llama.cpp/build/bin/llama-bench -m /models/qwen36-35b-a3b/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -t 8 -p 0 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -lm mlock -ctk q8_0 -ctv q8_0 -d 4096 --repack 1 --poll 50 -fa on"
    },
    {
      "config_id": "237b2d8a1c1acecc6f",
      "timestamp": "2026-10-07T00:31:07Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen36-35b-q4kxl",
      "model_name": "Qwen3.6-35B-A3B",
      "model_quant": "UD-Q4_K_XL",
      "runtime": "upstream",
      "phase": "context",
      "threads": "8",
      "threads_reported": "8",
      "build_commit": "0faee50",
      "backends": "CPU",
      "hf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
      "hf_revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
      "placement": "node1",
      "load_mode": "mlock",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "on",
      "fa_actual": "1",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "4096",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode_context",
      "mean_tok_s": "10.784284",
      "stddev_tok_s": "0.002294",
      "repetitions": "3",
      "model_bytes": "22349466112",
      "model_params": "34660610688",
      "peak_rss_gib": "21.289413452148438",
      "wall_s": "122.95674981101183",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27 --membind=1 /opt/llama.cpp/build/bin/llama-bench -m /models/qwen36-35b-a3b/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -t 8 -p 0 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -lm mlock -ctk f16 -ctv f16 -d 4096 --repack 1 --poll 50 -fa on"
    },
    {
      "config_id": "a9f60923c8ace7049f",
      "timestamp": "2026-10-07T00:35:35Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen36-35b-q4kxl",
      "model_name": "Qwen3.6-35B-A3B",
      "model_quant": "UD-Q4_K_XL",
      "runtime": "upstream",
      "phase": "context",
      "threads": "8",
      "threads_reported": "8",
      "build_commit": "0faee50",
      "backends": "CPU",
      "hf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
      "hf_revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
      "placement": "node1",
      "load_mode": "mlock",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "on",
      "fa_actual": "1",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "8192",
      "prompt_tokens": "0",
      "gen_tokens": "128",
      "test_type": "decode_context",
      "mean_tok_s": "9.667956",
      "stddev_tok_s": "0.005341",
      "repetitions": "3",
      "model_bytes": "22349466112",
      "model_params": "34660610688",
      "peak_rss_gib": "21.449966430664062",
      "wall_s": "191.55096793701523",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27 --membind=1 /opt/llama.cpp/build/bin/llama-bench -m /models/qwen36-35b-a3b/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -t 8 -p 0 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -lm mlock -ctk f16 -ctv f16 -d 8192 --repack 1 --poll 50 -fa on"
    }
  ],
  "batch": [
    {
      "config_id": "390089ddc8fa779278",
      "timestamp": "2026-10-07T00:16:28Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen36-35b-q4kxl",
      "model_name": "Qwen3.6-35B-A3B",
      "model_quant": "UD-Q4_K_XL",
      "runtime": "upstream",
      "phase": "batch",
      "threads": "8",
      "threads_reported": "8",
      "build_commit": "0faee50",
      "backends": "CPU",
      "hf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
      "hf_revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
      "placement": "node1",
      "load_mode": "mlock",
      "batch_size": "1024",
      "ubatch_size": "256",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "-1",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "512",
      "gen_tokens": "0",
      "test_type": "prefill",
      "mean_tok_s": "76.032487",
      "stddev_tok_s": "0.380385",
      "repetitions": "3",
      "model_bytes": "22349466112",
      "model_params": "34660610688",
      "peak_rss_gib": "21.038597106933594",
      "wall_s": "89.14904891900369",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27 --membind=1 /opt/llama.cpp/build/bin/llama-bench -m /models/qwen36-35b-a3b/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -t 8 -p 512 -n 128 -r 3 -ngl 0 -b 1024 -ub 256 -o json -lm mlock -ctk f16 -ctv f16 -d 0 --repack 1 --poll 50 -fa auto"
    },
    {
      "config_id": "47113082db237c51f0",
      "timestamp": "2026-10-07T00:13:26Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen36-35b-q4kxl",
      "model_name": "Qwen3.6-35B-A3B",
      "model_quant": "UD-Q4_K_XL",
      "runtime": "upstream",
      "phase": "batch",
      "threads": "8",
      "threads_reported": "8",
      "build_commit": "0faee50",
      "backends": "CPU",
      "hf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
      "hf_revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
      "placement": "node1",
      "load_mode": "mlock",
      "batch_size": "512",
      "ubatch_size": "128",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "-1",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "512",
      "gen_tokens": "0",
      "test_type": "prefill",
      "mean_tok_s": "70.468665",
      "stddev_tok_s": "0.154364",
      "repetitions": "3",
      "model_bytes": "22349466112",
      "model_params": "34660610688",
      "peak_rss_gib": "21.021690368652344",
      "wall_s": "91.73550409998279",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27 --membind=1 /opt/llama.cpp/build/bin/llama-bench -m /models/qwen36-35b-a3b/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -t 8 -p 512 -n 128 -r 3 -ngl 0 -b 512 -ub 128 -o json -lm mlock -ctk f16 -ctv f16 -d 0 --repack 1 --poll 50 -fa auto"
    },
    {
      "config_id": "d0b34373ae2c3452ab",
      "timestamp": "2026-10-07T00:14:58Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen36-35b-q4kxl",
      "model_name": "Qwen3.6-35B-A3B",
      "model_quant": "UD-Q4_K_XL",
      "runtime": "upstream",
      "phase": "batch",
      "threads": "8",
      "threads_reported": "8",
      "build_commit": "0faee50",
      "backends": "CPU",
      "hf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
      "hf_revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
      "placement": "node1",
      "load_mode": "mlock",
      "batch_size": "512",
      "ubatch_size": "256",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "-1",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "512",
      "gen_tokens": "0",
      "test_type": "prefill",
      "mean_tok_s": "75.130329",
      "stddev_tok_s": "0.755175",
      "repetitions": "3",
      "model_bytes": "22349466112",
      "model_params": "34660610688",
      "peak_rss_gib": "21.03802490234375",
      "wall_s": "89.77771489799488",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27 --membind=1 /opt/llama.cpp/build/bin/llama-bench -m /models/qwen36-35b-a3b/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -t 8 -p 512 -n 128 -r 3 -ngl 0 -b 512 -ub 256 -o json -lm mlock -ctk f16 -ctv f16 -d 0 --repack 1 --poll 50 -fa auto"
    },
    {
      "config_id": "d08861a39c701a9206",
      "timestamp": "2026-10-07T00:19:29Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen36-35b-q4kxl",
      "model_name": "Qwen3.6-35B-A3B",
      "model_quant": "UD-Q4_K_XL",
      "runtime": "upstream",
      "phase": "batch",
      "threads": "8",
      "threads_reported": "8",
      "build_commit": "0faee50",
      "backends": "CPU",
      "hf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
      "hf_revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
      "placement": "node1",
      "load_mode": "mlock",
      "batch_size": "2048",
      "ubatch_size": "1024",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "-1",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "512",
      "gen_tokens": "0",
      "test_type": "prefill",
      "mean_tok_s": "76.026498",
      "stddev_tok_s": "0.249196",
      "repetitions": "3",
      "model_bytes": "22349466112",
      "model_params": "34660610688",
      "peak_rss_gib": "21.063949584960938",
      "wall_s": "89.57586365999305",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27 --membind=1 /opt/llama.cpp/build/bin/llama-bench -m /models/qwen36-35b-a3b/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -t 8 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 1024 -o json -lm mlock -ctk f16 -ctv f16 -d 0 --repack 1 --poll 50 -fa auto"
    },
    {
      "config_id": "daf2d5eef4ffb4170e",
      "timestamp": "2026-10-07T00:17:58Z",
      "kernel_release": "7.0.14-20-pve",
      "model_id": "qwen36-35b-q4kxl",
      "model_name": "Qwen3.6-35B-A3B",
      "model_quant": "UD-Q4_K_XL",
      "runtime": "upstream",
      "phase": "batch",
      "threads": "8",
      "threads_reported": "8",
      "build_commit": "0faee50",
      "backends": "CPU",
      "hf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
      "hf_revision": "a483e9e6cbd595906af30beda3187c2663a1118c",
      "placement": "node1",
      "load_mode": "mlock",
      "batch_size": "2048",
      "ubatch_size": "512",
      "kv_k": "f16",
      "kv_v": "f16",
      "fa_requested": "auto",
      "fa_actual": "-1",
      "repack_requested": "1",
      "repack_actual": "True",
      "depth_tokens": "0",
      "prompt_tokens": "512",
      "gen_tokens": "0",
      "test_type": "prefill",
      "mean_tok_s": "75.105146",
      "stddev_tok_s": "0.525857",
      "repetitions": "3",
      "model_bytes": "22349466112",
      "model_params": "34660610688",
      "peak_rss_gib": "21.06397247314453",
      "wall_s": "89.96021015397855",
      "protocol_version": "2",
      "status": "ok",
      "error": "",
      "command": "numactl --physcpubind=20,21,22,23,24,25,26,27 --membind=1 /opt/llama.cpp/build/bin/llama-bench -m /models/qwen36-35b-a3b/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf -t 8 -p 512 -n 128 -r 3 -ngl 0 -b 2048 -ub 512 -o json -lm mlock -ctk f16 -ctv f16 -d 0 --repack 1 --poll 50 -fa auto"
    }
  ],
  "concurrency": [
    {
      "C": 1,
      "batches": 6,
      "tasks_done": 6,
      "wall_s_total": 49.13154201301222,
      "tasks_per_hour": 439.6361098188076,
      "aggregate_tok_s": 12.41564939766077,
      "total_tokens": 610,
      "latency_p95_s": 13.677993661000073
    },
    {
      "C": 2,
      "batches": 6,
      "tasks_done": 12,
      "wall_s_total": 58.57107381199603,
      "tasks_per_hour": 737.5654429465512,
      "aggregate_tok_s": 22.96355372136832,
      "total_tokens": 1345,
      "latency_p95_s": 14.098483088202917
    },
    {
      "C": 4,
      "batches": 6,
      "tasks_done": 24,
      "wall_s_total": 292.9826062570064,
      "tasks_per_hour": 294.89805249465667,
      "aggregate_tok_s": 11.184964328992939,
      "total_tokens": 3277,
      "latency_p95_s": 104.46766705514666
    }
  ]
}