{ "amendments": [ { "id": "A-L01", "new": "swap used < 80 % AND MemAvailable >= 10 GiB AND PSI memory full avg60 < 10 % (/dev/shm < 4 GiB kept)", "old": "swap used < 50 %, /dev/shm used < 4 GiB, MemAvailable >= 10 GiB", "readings_at_1824Z": { "MemAvailable_GiB": 49.2, "MemFree_GiB": 41.0, "bench_unit_MemorySwapCurrent_bytes": 2994176, "largest_swap_holders": "two stale sharpsignal.memory_mcp processes (pids 9106, 12122; ~1.1 GB each) in old ghostty scopes, and the system ollama's nomic runner (0.34 GB)", "psi_memory_full_avg60_pct": 1.74, "swap_img": "0.3 of 2.0 GiB", "zram0": "9.9 of 16.0 GiB used, stored in 3.56 GiB of RAM" }, "reason": "the old gate tripped at 18:24:03Z (dry12-anchor-4096/REFUSED.json: swap 54.2 %) with no memory risk: this box's swap is mostly zram (compressed RAM), so a bare percentage over-reads the risk; the new gate still catches the 15:56Z case (swap 17 of 17 GiB)", "rows_the_author_had_seen": "none scored: the gate tripped before item 12's first load", "ruled_by": "the lead under R-0 (message to the laptop lane, received ~18:25Z; applied 18:25:59Z)", "utc": "2026-09-23T18:25Z", "what": "the memory gate (PLAN-v2 §7, G-2)" }, { "id": "A-L02", "new": "MemoryMax=16G", "old": "MemoryMax=4G", "reason": "item 14 measured: under 4G the serve unit hit its cap 633,419 times (memory.events max) during item 12, swapped up to 6.0 GB (MemorySwapPeak 6,008,573,952 B), memory.pressure full avg60 read 9.41 %, and prefill stalled 11,416 ms on doorman item A5 and 4,607 ms on C1 (515-521 prompt tokens; the other items 121-342 ms). No OOM, but the cap throttled the measurement. 16G is a ceiling that does not bind: llama-server's host prompt cache is bounded (~8 GiB) and the rest is clean page cache; each arm.json records memory.events, memory.stat and memory.pressure before and after, so a binding cap is visible", "rows_the_author_had_seen": "dry12-anchor-4096 and dry12-anchor-131072 under 4G (kept as superseded receipts)", "ruled_by": "the lane, under PLAN-v2 Appendix A/B item 14 ('raise to a measured value if it OOMs; never remove')", "utc": "2026-09-23T18:31Z", "what": "the bench instance's MemoryMax (Appendix A: 4G, marked UNTESTED)" }, { "id": "A-L03", "new": "MemoryMax=32G (applied live: systemctl --user set-property)", "old": "MemoryMax=16G (A-L02)", "reason": "16G still bound: llama-server's host prompt cache grows to its own 8,192 MiB limit (its log: '54 prompts, 8139.974 MiB (limits: 8192.000 MiB ...)'), and with an ~18 GB model file's page cache charged to the same cgroup the cap was hit 5,598 times during k0a-gemma4-anchor-cd and 6,994 times during b-vicuna-33b; both arms show 1.4-7.2 s prefill stalls on 1-3 items that benchbox's same arms do not. Anon stays <= ~9 GB; the rest is reclaimable page cache, so 32G is a ceiling that should not bind, and every arm.json carries the cgroup's memory.events before/after to prove it", "rows_the_author_had_seen": "k0a-gemma4-anchor-cd, k1-mistral-v0.1-cd, b-falcon-7b, b-vicuna-33b, and b-llama2-7b's task c", "ruled_by": "the lane, under PLAN-v2 Appendix B item 14 (as A-L02)", "utc": "2026-09-23T18:51:52Z", "what": "the bench instance's MemoryMax" }, { "id": "A-L04", "new": "nvidia-powerd active, clocks.max.sm 3090 (lock off), AC online, one board, and enforced.power.limit anywhere in the Dynamic Boost envelope (> 95 W and <= 175 W); the limit is reported per task as min/median/max from the 1 Hz ext trace, never as one number", "old": "enforced.power.limit == 150.00 W (+-0.5), clocks.max.sm 3090, powerd active, AC on, one board", "reason": "under Dynamic Boost the firmware moves the GPU's enforced limit when the CPU draws: the 1 Hz traces read 135.00-150.00 W on b-falcon-7b and b-vicuna-33b and 137.08-150.00 W on b-llama2-7b (a steady 150.00 on the anchor and mistral; loadavg ~5 from peers); the old gate refused b-llama2-7b before its task d at 18:51:55Z (146.76 W). Memory laptop-5090-refuses-a-power-cap: the boost posture's cap is 'a range from the trace'", "rows_the_author_had_seen": "the same as A-L03", "ruled_by": "the lead under R-0 (message to the laptop lane, received 18:53Z)", "utc": "2026-09-23T18:53Z", "what": "the laptop posture gate (PLAN-v2 §2.2: 'enforced.power.limit 150.00 W ... gated per arm, rc 4 on mismatch')" } ], "argv": [ "coilaptop.py", "--arm-id", "b-gemma4-31b", "--out", "/workshop/estate/bench/cost-of-intelligence-2026-09-23/laptop-reading-0/rows/b-gemma4-31b", "--model", "gemma4:31b-it-q4_K_M", "--num-ctx", "4096", "--think", "off", "--think-budget", "4096", "--tasks", "c,d", "--probes" ], "arm_id": "b-gemma4-31b", "baseline_rows": [ { "memory_mib": 15.0, "sm_mhz": 1192.0, "temp_c": 34.0, "uuid": "", "watts": 11.27 } ], "board": "", "box": "the laptop", "box_class": "gpu-5090-laptop-24g", "digest_pin": "6316f0629137", "disk_floor": { "df": "Avail\n111286697984", "floor_gib": 60.0, "free_gib": 103.64381408691406, "store": "/workshop/bench-coi-2026-09-23/store" }, "ended": "2026-09-23T21:30:23Z", "energy_instruments": [ "board (nvidia-smi power.draw by UUID, 1 Hz; run.py Boards)", "rapl psys + package-0 (+ core, uncore) via /run/rapl/energy (R-10d); not wall" ], "host": "http://127.0.0.1:withheld-2026-10-14", "idle_preload": { "board": { "cards": [ "" ], "joules": 157.6438596534729, "mean_watts": 11.25464838987203, "peak_watts": 11.32, "per_card": { "": { "joules": 157.6438596534729, "mean_watts": 11.25464838987203, "min_watts": 11.19, "peak_watts": 11.32, "samples": 15, "seconds": 14.006999969482422 } }, "samples": 15, "seconds": 14.006999969482422 }, "rapl": { "domains": { "core": { "joules": 161.037735, "mean_watts": 13.033376898022905, "samples": 13, "seconds": 12.35579514503479, "wraps": 0 }, "package-0": { "joules": 274.592803, "mean_watts": 22.22374452526228, "samples": 13, "seconds": 12.35582971572876, "wraps": 0 }, "psys": { "joules": 3.190422, "mean_watts": 0.25821392620283057, "samples": 13, "seconds": 12.355731725692749, "wraps": 0 }, "uncore": { "joules": 0.329284, "mean_watts": 0.026649824705652483, "samples": 13, "seconds": 12.355953693389893, "wraps": 0 } }, "instrument": "RAPL via /run/rapl/energy (root sampler, 1 Hz; R-10d): psys = the platform's own reading (not wall); package-0 = the CPU package" }, "window_epoch": [ 1790198778.418, 1790198792.425 ] }, "instrument": { "coilaptop_py": "0b5dcb41ac989a57dfe4912740dfd24ccd600da5ba3fbc181688d18a8adf94f0", "jev/kit/doorman_planted.json": "4daf059c26a47a84fa3b73630a64ba1a348da72c2e5df7f3c28c007b67da22f7", "jev/kit/task_c.json": "61a0b1d6e7cf66fd91992c0c3a4572f6fbae90e6471048fd7dc1eb41ff2d7f39", "jev/run.py": "withheld-2026-10-14", "jev/run_addenda.py": "a34aca61833ed0c5860c91c7e2e227429e8ea9e2e50fec68d7b7fec15d89acab" }, "limit": null, "memory_gate": { "amendment": "A-L01", "limits": { "mem_available_gib_min": 10.0, "psi_full_avg60_max_pct": 10.0, "shm_used_gib_max": 4.0, "swap_used_frac_max": 0.8 }, "mem_available_gib": 42.20608139038086, "psi_memory": { "full": { "avg10": 0.0, "avg300": 0.6, "avg60": 0.0, "total": 118569978.0 }, "some": { "avg10": 0.0, "avg300": 0.63, "avg60": 0.0, "total": 126460131.0 } }, "shm_used_gib": 3.1962127685546875, "stamp": "2026-09-23T21:26:12Z", "swap_total_gib": 17.99999237060547, "swap_used_frac": 0.3428, "swaps": [ "Filename\t\t\t\tType\t\tSize\t\tUsed\t\tPriority", "/swap.img file\t\t2097148\t\t59120\t\t-1", "/dev/zram0 partition\t16777212\t6411828\t\t100" ], "why": "ok" }, "model": "gemma4:31b-it-q4_K_M", "num_ctx": 4096, "ollama_version": "0.32.13", "pin": { "api_ps_model": { "context_length": 4096, "details": { "families": [ "gemma4" ], "family": "gemma4", "format": "gguf", "parameter_size": "31.3B", "parent_model": "", "quantization_level": "Q4_K_M" }, "digest": "6316f0629137b426c9d9b853ffc4c8209589f30ee39aebede6285096c0ff47e7", "expires_at": "2319-01-03T21:13:55.30574498Z", "model": "gemma4:31b-it-q4_K_M", "name": "gemma4:31b-it-q4_K_M", "size": 20253381754, "size_vram": 20253381754 }, "card": "", "cards": [ "" ], "fit": { "growth_bytes": 22120759296.0, "ps_size": 20253381754, "ps_size_vram": 20253381754, "ps_whole": true, "rule": "WHOLE when the named board grew by at least 90 %% of the store's bytes for this tag (weights + projector/draft; KV rides on top) AND /api/ps does not report a CPU share; ps alone is corroboration, never the verdict", "store_bytes": 19868981791, "whole": true }, "floor_mib": 200.0, "growth_mib": { "": 21096.0 }, "memory_after_mib": { "": 21111.0 }, "named_grew": true, "ok": true, "others_quiet": true, "per_card_grew": { "": true } }, "postflight": { "after_unload": { "api_ps": { "models": [] }, "api_ps_empty": true, "back_to_baseline": true, "board_mib_baseline": 15.0, "board_mib_now": 15.0, "rows": [ { "memory_mib": 15.0, "sm_mhz": 1192.0, "temp_c": 47.0, "uuid": "", "watts": 16.63 } ], "stamp": "2026-09-23T21:30:23Z" }, "posture_after": { "ac_online": "1", "battery": { "capacity": "93", "status": "Not charging" }, "boards_on_bus": [ "0, , NVIDIA GeForce RTX 5090 Laptop GPU, withheld-2026-10-14" ], "card": { "clocks.max.sm": "3090", "clocks.mem": "14001", "clocks.sm": "1852", "clocks_throttle_reasons.active": "0x0000000000000004", "driver_version": "595.91.07", "enforced.power.limit": "150.00", "memory.total": "24463", "memory.used": "21121", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "pci.bus_id": "withheld-2026-10-14", "pcie.link.gen.current": "5", "pcie.link.gen.max": "5", "pcie.link.width.current": "8", "pcie.link.width.max": "16", "persistence_mode": "Enabled", "power.default_limit": "95.00", "power.draw": "93.15", "power.limit": "[N/A]", "power.max_limit": "175.00", "pstate": "P0", "temperature.gpu": "55", "utilization.gpu": "89", "uuid": "", "vbios_version": "98.03.A2.00.21" }, "compute_apps": [ ", 137540, /workshop/bench-coi-2026-09-23/ollama-0.32.13/lib/ollama/llama-server, 21098 MiB" ], "fan": "unreadable on this board (fan.speed N/A); the operator's turbo fan per OPERATOR-CONDITIONS.md", "how": "nvidia-smi -i ; enforced.power.limit and clocks.max.sm READ, never typed; power.limit reads N/A on this board and the board refuses -pl", "loadavg": "3.27 2.80 2.73 4/1909 171767", "nvidia_powerd": "active", "stamp": "2026-09-23T21:30:17Z" }, "seat_after": { "api_ps": { "models": [] }, "stamp": "2026-09-23T21:30:23Z" }, "serve_cgroup_memory_after": { "events": { "high": "0", "low": "0", "max": "43078", "oom": "0", "oom_group_kill": "0", "oom_kill": "0", "sock_throttled": "401" }, "pressure": [ "some avg10=0.00 avg60=0.00 avg300=0.00 total=15976543", "full avg10=0.00 avg60=0.00 avg300=0.00 total=15976543" ], "stat": { "anon": 61198336, "file": 17378217984, "file_mapped": 1757184, "shmem": 0 }, "swap_current": 2621440 }, "serve_unit_after": { "ActiveState": "active", "ControlGroup": "/user.slice/user-1000.slice/user@1000.service/app.slice/coi-ollama-laptop.service", "Environment": "TZ=UTC OLLAMA_HOST=127.0.0.1:withheld-2026-10-14 OLLAMA_KEEP_ALIVE=-1 OLLAMA_MAX_LOADED_MODELS=1 OLLAMA_=1 OLLAMA_FLASH_ATTENTION=1 OLLAMA_KV_CACHE_TYPE=f16 OLLAMA_NOPRUNE=1 OLLAMA_MODELS=/workshop/bench-coi-2026-09-23/store CUDA_VISIBLE_DEVICES= OLLAMA_VULKAN=0 GGML_VK_VISIBLE_DEVICES=-1 HIP_VISIBLE_DEVICES= ROCR_VISIBLE_DEVICES=", "ExecStart": "{ path=/workshop/bench-coi-2026-09-23/ollama-0.32.13/bin/ollama ; argv[]=/workshop/bench-coi-2026-09-23/ollama-0.32.13/bin/ollama serve ; ignore_errors=no ; start_time=[n/a] ; stop_time=[n/a] ; pid=0 ; code=(null) ; status=0/0 }", "MainPID": "1739347", "MemoryCurrent": "17943482368", "MemoryMax": "34359738368", "MemoryPeak": "30671613952", "NRestarts": "0" } }, "posture_before": { "ac_online": "1", "battery": { "capacity": "93", "status": "Not charging" }, "boards_on_bus": [ "0, , NVIDIA GeForce RTX 5090 Laptop GPU, withheld-2026-10-14" ], "card": { "clocks.max.sm": "3090", "clocks.mem": "810", "clocks.sm": "1192", "clocks_throttle_reasons.active": "0x0000000000000000", "driver_version": "595.91.07", "enforced.power.limit": "150.00", "memory.total": "24463", "memory.used": "15", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "pci.bus_id": "withheld-2026-10-14", "pcie.link.gen.current": "2", "pcie.link.gen.max": "5", "pcie.link.width.current": "8", "pcie.link.width.max": "16", "persistence_mode": "Enabled", "power.default_limit": "95.00", "power.draw": "11.32", "power.limit": "[N/A]", "power.max_limit": "175.00", "pstate": "P5", "temperature.gpu": "34", "utilization.gpu": "0", "uuid": "", "vbios_version": "98.03.A2.00.21" }, "compute_apps": [], "fan": "unreadable on this board (fan.speed N/A); the operator's turbo fan per OPERATOR-CONDITIONS.md", "how": "nvidia-smi -i ; enforced.power.limit and clocks.max.sm READ, never typed; power.limit reads N/A on this board and the board refuses -pl", "loadavg": "2.00 2.52 2.66 2/1878 132845", "nvidia_powerd": "active", "stamp": "2026-09-23T21:26:12Z" }, "posture_verdict": "Dynamic Boost: enforced 150.00 W (inside the (95, 175] W envelope), clock lock off (3090 MHz), powerd active, AC on, one board", "probes": { "decode": { "board": { "cards": [ "" ], "joules": 3333.6605321943757, "mean_watts": 128.17827381560792, "peak_watts": 139.14, "per_card": { "": { "joules": 3333.6605321943757, "mean_watts": 128.17827381560792, "min_watts": 43.29, "peak_watts": 139.14, "samples": 27, "seconds": 26.007999897003174 } }, "samples": 27, "seconds": 26.007999897003174 }, "num_predict": 256, "prompt": "Describe, in about three hundred words, a lighthouse on a rocky coast during a winter storm.", "rapl": { "domains": { "core": { "joules": 726.179232, "mean_watts": 29.347291244088893, "samples": 25, "seconds": 24.744335889816284, "wraps": 0 }, "package-0": { "joules": 933.795034, "mean_watts": 37.73772249143083, "samples": 25, "seconds": 24.74433994293213, "wraps": 0 }, "psys": { "joules": 6.740644, "mean_watts": 0.2724057494998, "samples": 25, "seconds": 24.744866847991943, "wraps": 0 }, "uncore": { "joules": 0.680906, "mean_watts": 0.02751668171201661, "samples": 25, "seconds": 24.74520754814148, "wraps": 0 } }, "instrument": "RAPL via /run/rapl/energy (root sampler, 1 Hz; R-10d): psys = the platform's own reading (not wall); package-0 = the CPU package" }, "runs": [ { "decode_tok_s": 32.47251569399221, "done_reason": "length", "eval_count": 256, "eval_s": 7.88359, "load_s": 0.238174134, "prefill_tok_s": 336.9378665515462, "prompt_eval_count": 32, "prompt_eval_s": 0.094973, "runner_pids": [ 137540, 1739347 ], "text_head": "The lighthouse stands as a solitary sentinel upon a jagged pedestal of basalt, clinging to the edge of a coast that seem", "thinking_chars": 0, "total_s": 8.548899166 }, { "decode_tok_s": 32.47812262242064, "done_reason": "length", "eval_count": 256, "eval_s": 7.882229, "load_s": 0.261615309, "prefill_tok_s": 421.7185028993147, "prompt_eval_count": 32, "prompt_eval_s": 0.07588, "runner_pids": [ 137540, 1739347 ], "text_head": "The lighthouse stands as a solitary sentinel upon a jagged pedestal of basalt, clinging to the edge of a coast that seem", "thinking_chars": 0, "total_s": 8.331235889 }, { "decode_tok_s": 32.28059096683144, "done_reason": "length", "eval_count": 256, "eval_s": 7.930462, "load_s": 0.266416346, "prefill_tok_s": 429.3803504817111, "prompt_eval_count": 32, "prompt_eval_s": 0.074526, "runner_pids": [ 137540, 1739347 ], "text_head": "The lighthouse stands as a solitary sentinel upon a jagged pedestal of basalt, clinging to the edge of a coast that seem", "thinking_chars": 0, "total_s": 8.3834422 } ], "window_epoch": [ 1790198965.358, 1790198991.366 ] }, "prefill": { "board": { "cards": [ "" ], "joules": 1140.9977131271362, "mean_watts": 87.72182023558035, "peak_watts": 150.86, "per_card": { "": { "joules": 1140.9977131271362, "mean_watts": 87.72182023558035, "min_watts": 24.27, "peak_watts": 150.86, "samples": 14, "seconds": 13.006999969482422 } }, "samples": 14, "seconds": 13.006999969482422 }, "rapl": { "domains": { "core": { "joules": 318.457742, "mean_watts": 25.74835296228431, "samples": 13, "seconds": 12.368082046508789, "wraps": 0 }, "package-0": { "joules": 426.845648, "mean_watts": 34.51234901287367, "samples": 13, "seconds": 12.367910623550415, "wraps": 0 }, "psys": { "joules": 3.425589, "mean_watts": 0.27696997822973146, "samples": 13, "seconds": 12.368087768554688, "wraps": 0 }, "uncore": { "joules": 0.343078, "mean_watts": 0.027738368261647155, "samples": 13, "seconds": 12.368355512619019, "wraps": 0 } }, "instrument": "RAPL via /run/rapl/energy (root sampler, 1 Hz; R-10d): psys = the platform's own reading (not wall); package-0 = the CPU package" }, "runs": [ { "decode_tok_s": 1000000.0, "done_reason": "length", "eval_count": 1, "eval_s": 1e-06, "load_s": 0.237846962, "prefill_tok_s": 1095.6324242051853, "prompt_eval_count": 2220, "prompt_eval_s": 2.026227, "runner_pids": [ 137540, 1739347 ], "text_head": "Darwin", "thinking_chars": 0, "total_s": 2.288134512 }, { "decode_tok_s": 1000000.0, "done_reason": "length", "eval_count": 1, "eval_s": 1e-06, "load_s": 0.251101026, "prefill_tok_s": 1068.8909370139345, "prompt_eval_count": 2218, "prompt_eval_s": 2.075047999, "runner_pids": [ 137540, 1739347 ], "text_head": "Darwin", "thinking_chars": 0, "total_s": 3.296473572 }, { "decode_tok_s": 1000000.0, "done_reason": "length", "eval_count": 1, "eval_s": 1e-06, "load_s": 0.256610446, "prefill_tok_s": 1066.977223746892, "prompt_eval_count": 2219, "prompt_eval_s": 2.079706999, "runner_pids": [ 137540, 1739347 ], "text_head": "Darwin", "thinking_chars": 0, "total_s": 3.346384688 }, { "decode_tok_s": 1000000.0, "done_reason": "length", "eval_count": 1, "eval_s": 1e-06, "load_s": 0.23469321, "prefill_tok_s": 1064.9519883345024, "prompt_eval_count": 2218, "prompt_eval_s": 2.082723, "runner_pids": [ 137540, 1739347 ], "text_head": "Darwin", "thinking_chars": 0, "total_s": 3.295418034 } ], "target_tokens_approx": 3000.0, "window_epoch": [ 1790198992.424, 1790199005.431 ] }, "ttft": { "how": "the first 12 task_c prompts, run.py's body with stream true, in task order (so the prompt-cache condition is the task's own); TTFT = client clock on localhost from the request to the first chunk carrying content or thinking", "median_s": 0.7770810354995774, "p95_s": 0.8310550629976206, "runs": [ { "eval_count": 4, "prompt_eval_count": 225, "server_load_ms": 277.038307, "server_prompt_eval_ms": 260.801, "total_s": 1.5699196530040354, "ttft_s": 1.472671622002963 }, { "eval_count": 4, "prompt_eval_count": 322, "server_load_ms": 238.439148, "server_prompt_eval_ms": 292.645, "total_s": 0.7734512359893415, "ttft_s": 0.6768386799958535 }, { "eval_count": 4, "prompt_eval_count": 241, "server_load_ms": 256.017843, "server_prompt_eval_ms": 231.597, "total_s": 0.875308547008899, "ttft_s": 0.778281884995522 }, { "eval_count": 4, "prompt_eval_count": 342, "server_load_ms": 235.954907, "server_prompt_eval_ms": 317.467, "total_s": 0.8121174289990449, "ttft_s": 0.7155463700037217 }, { "eval_count": 4, "prompt_eval_count": 349, "server_load_ms": 245.470153, "server_prompt_eval_ms": 319.288, "total_s": 0.8782416629983345, "ttft_s": 0.7835291009978391 }, { "eval_count": 4, "prompt_eval_count": 351, "server_load_ms": 228.919895, "server_prompt_eval_ms": 321.632, "total_s": 0.8738030420063296, "ttft_s": 0.7758801860036328 }, { "eval_count": 4, "prompt_eval_count": 261, "server_load_ms": 252.054526, "server_prompt_eval_ms": 239.061, "total_s": 0.8039071720122593, "ttft_s": 0.7078047510003671 }, { "eval_count": 4, "prompt_eval_count": 241, "server_load_ms": 249.848636, "server_prompt_eval_ms": 232.806, "total_s": 0.8247735690092668, "ttft_s": 0.7279184970102506 }, { "eval_count": 4, "prompt_eval_count": 312, "server_load_ms": 253.736305, "server_prompt_eval_ms": 292.659, "total_s": 0.8024362010037294, "ttft_s": 0.7050502329948358 }, { "eval_count": 4, "prompt_eval_count": 350, "server_load_ms": 275.827861, "server_prompt_eval_ms": 321.868, "total_s": 0.8944845429941779, "ttft_s": 0.7955501139949774 }, { "eval_count": 4, "prompt_eval_count": 342, "server_load_ms": 269.527938, "server_prompt_eval_ms": 323.414, "total_s": 0.9125757459987653, "ttft_s": 0.8078997770062415 }, { "eval_count": 4, "prompt_eval_count": 342, "server_load_ms": 298.999098, "server_prompt_eval_ms": 321.997, "total_s": 0.926254058998893, "ttft_s": 0.8310550629976206 } ] } }, "rapl_probe": [ "1790198772.203635348 package-0 91249452619 262143328850", "1790198772.211399698 core 202502174586 262143328850", "1790198772.218923331 uncore 7664648469 262143328850", "1790198772.225739590 psys 22755737643 262143328850" ], "rapl_unreadable_polls": 0, "seat_before": { "api_ps": { "models": [] }, "stamp": "2026-09-23T21:26:12Z" }, "serve_cgroup_memory_before": { "events": { "high": "0", "low": "0", "max": "43078", "oom": "0", "oom_group_kill": "0", "oom_kill": "0", "sock_throttled": "401" }, "pressure": [ "some avg10=0.00 avg60=0.00 avg300=0.00 total=15976543", "full avg10=0.00 avg60=0.00 avg300=0.00 total=15976543" ], "stat": { "anon": 31858688, "file": 20618612736, "file_mapped": 1757184, "shmem": 0 }, "swap_current": 2629632 }, "serve_unit_before": { "ActiveState": "active", "ControlGroup": "/user.slice/user-1000.slice/user@1000.service/app.slice/coi-ollama-laptop.service", "Environment": "TZ=UTC OLLAMA_HOST=127.0.0.1:withheld-2026-10-14 OLLAMA_KEEP_ALIVE=-1 OLLAMA_MAX_LOADED_MODELS=1 OLLAMA_=1 OLLAMA_FLASH_ATTENTION=1 OLLAMA_KV_CACHE_TYPE=f16 OLLAMA_NOPRUNE=1 OLLAMA_MODELS=/workshop/bench-coi-2026-09-23/store CUDA_VISIBLE_DEVICES= OLLAMA_VULKAN=0 GGML_VK_VISIBLE_DEVICES=-1 HIP_VISIBLE_DEVICES= ROCR_VISIBLE_DEVICES=", "ExecStart": "{ path=/workshop/bench-coi-2026-09-23/ollama-0.32.13/bin/ollama ; argv[]=/workshop/bench-coi-2026-09-23/ollama-0.32.13/bin/ollama serve ; ignore_errors=no ; start_time=[n/a] ; stop_time=[n/a] ; pid=0 ; code=(null) ; status=0/0 }", "MainPID": "1739347", "MemoryCurrent": "21243908096", "MemoryMax": "34359738368", "MemoryPeak": "30671613952", "NRestarts": "0" }, "show": { "capabilities": [ "completion", "vision", "tools", "thinking" ], "details": { "families": [ "gemma4" ], "family": "gemma4", "format": "gguf", "parameter_size": "31.3B", "parent_model": "", "quantization_level": "Q4_K_M" }, "model_info": { "gemma4.attention.head_count": 32, "gemma4.attention.head_count_kv": null, "gemma4.attention.key_length": 512, "gemma4.attention.key_length_swa": 256, "gemma4.attention.layer_norm_rms_epsilon": 1e-06, "gemma4.attention.shared_kv_layers": 0, "gemma4.attention.sliding_window": 1024, "gemma4.attention.sliding_window_pattern": null, "gemma4.attention.value_length": 512, "gemma4.attention.value_length_swa": 256, "gemma4.block_count": 60, "gemma4.context_length": 262144, "gemma4.embedding_length": 5376, "gemma4.embedding_length_per_layer_input": 0, "gemma4.feed_forward_length": 21504, "gemma4.final_logit_softcapping": 30, "gemma4.rope.dimension_count": 512, "gemma4.rope.dimension_count_swa": 256, "gemma4.rope.freq_base": 1000000, "gemma4.rope.freq_base_swa": 10000, "gemma4.vision.attention.head_count": 16, "gemma4.vision.attention.layer_norm_epsilon": 1e-06, "gemma4.vision.block_count": 27, "gemma4.vision.embedding_length": 1152, "gemma4.vision.feed_forward_length": 4304, "gemma4.vision.num_channels": 3, "gemma4.vision.patch_size": 16, "gemma4.vision.projector.scale_factor": 3, "general.architecture": "gemma4", "general.file_type": 15, "general.parameter_count": 31273089132, "general.quantization_version": 2, "tokenizer.ggml.add_bos_token": false, "tokenizer.ggml.add_eos_token": false, "tokenizer.ggml.add_mask_token": false, "tokenizer.ggml.add_padding_token": false, "tokenizer.ggml.add_unknown_token": false, "tokenizer.ggml.bos_token_id": 2, "tokenizer.ggml.eos_token_id": 1, "tokenizer.ggml.eos_token_ids": null, "tokenizer.ggml.mask_token_id": 4, "tokenizer.ggml.merges": null, "tokenizer.ggml.model": "llama", "tokenizer.ggml.padding_token_id": 0, "tokenizer.ggml.pre": "gemma4", "tokenizer.ggml.scores": null, "tokenizer.ggml.token_type": null, "tokenizer.ggml.tokens": null, "tokenizer.ggml.unknown_token_id": 3 }, "parameters": "top_k 64\ntop_p 0.95\ntemperature 1", "system_sha256": null, "template_head": "{{ .Prompt }}", "template_sha256": "b507b9c2f6ca642bffcd06665ea7c91f235fd32daeefdf875a0f938db05fb315" }, "started": "2026-09-23T21:26:12Z", "tag": { "details": { "families": [ "gemma4" ], "family": "gemma4", "format": "gguf", "parameter_size": "31.3B", "parent_model": "", "quantization_level": "Q4_K_M" }, "digest": "6316f0629137b426c9d9b853ffc4c8209589f30ee39aebede6285096c0ff47e7", "modified_at": "2026-09-23T21:26:12.466787004Z", "name": "gemma4:31b-it-q4_K_M", "size": 19868981791 }, "tasks": [ "c", "d" ], "tasks_run": { "c": { "board_idle_window_epoch": [ 1790198801.57, 1790198810.575 ], "board_task_window_epoch": [ 1790198811.603, 1790198898.648 ], "energy": { "cards": [ "" ], "decisions": 108, "joules": 7260.833930127621, "joules_per_decision": 67.22994379747797, "mean_watts": 83.41471565010723, "net_joules_per_decision": 48.928103774905786, "peak_watts": 110.11, "per_card": { "": { "joules": 7260.833930127621, "mean_watts": 83.41471565010723, "min_watts": 22.67, "peak_watts": 110.11, "samples": 88, "seconds": 87.04500007629395 } }, "samples": 88, "seconds": 87.04500007629395 }, "idle": { "cards": [ "" ], "joules": 204.48356259584426, "mean_watts": 22.70778012183732, "peak_watts": 22.75, "per_card": { "": { "joules": 204.48356259584426, "mean_watts": 22.70778012183732, "min_watts": 22.65, "peak_watts": 22.75, "samples": 10, "seconds": 9.005000114440918 } }, "samples": 10, "seconds": 9.005000114440918, "what": "the draw of these cards immediately before the arm, with the model resident and no work in flight" }, "posture_end": { "ac_online": "1", "battery": { "capacity": "93", "status": "Not charging" }, "boards_on_bus": [ "0, , NVIDIA GeForce RTX 5090 Laptop GPU, withheld-2026-10-14" ], "card": { "clocks.max.sm": "3090", "clocks.mem": "14001", "clocks.sm": "1590", "clocks_throttle_reasons.active": "0x0000000000000000", "driver_version": "595.91.07", "enforced.power.limit": "150.00", "memory.total": "24463", "memory.used": "21117", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "pci.bus_id": "withheld-2026-10-14", "pcie.link.gen.current": "5", "pcie.link.gen.max": "5", "pcie.link.width.current": "8", "pcie.link.width.max": "16", "persistence_mode": "Enabled", "power.default_limit": "95.00", "power.draw": "43.44", "power.limit": "[N/A]", "power.max_limit": "175.00", "pstate": "P0", "temperature.gpu": "45", "utilization.gpu": "0", "uuid": "", "vbios_version": "98.03.A2.00.21" }, "compute_apps": [ ", 137540, /workshop/bench-coi-2026-09-23/ollama-0.32.13/lib/ollama/llama-server, 21094 MiB" ], "fan": "unreadable on this board (fan.speed N/A); the operator's turbo fan per OPERATOR-CONDITIONS.md", "how": "nvidia-smi -i ; enforced.power.limit and clocks.max.sm READ, never typed; power.limit reads N/A on this board and the board refuses -pl", "loadavg": "2.91 2.64 2.68 2/1886 149309", "nvidia_powerd": "active", "stamp": "2026-09-23T21:28:20Z" }, "posture_start": { "ac_online": "1", "battery": { "capacity": "93", "status": "Not charging" }, "boards_on_bus": [ "0, , NVIDIA GeForce RTX 5090 Laptop GPU, withheld-2026-10-14" ], "card": { "clocks.max.sm": "3090", "clocks.mem": "14001", "clocks.sm": "1590", "clocks_throttle_reasons.active": "0x0000000000000000", "driver_version": "595.91.07", "enforced.power.limit": "150.00", "memory.total": "24463", "memory.used": "21111", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "pci.bus_id": "withheld-2026-10-14", "pcie.link.gen.current": "5", "pcie.link.gen.max": "5", "pcie.link.width.current": "8", "pcie.link.width.max": "16", "persistence_mode": "Enabled", "power.default_limit": "95.00", "power.draw": "22.65", "power.limit": "[N/A]", "power.max_limit": "175.00", "pstate": "P0", "temperature.gpu": "35", "utilization.gpu": "0", "uuid": "", "vbios_version": "98.03.A2.00.21" }, "compute_apps": [ ", 137540, /workshop/bench-coi-2026-09-23/ollama-0.32.13/lib/ollama/llama-server, 21088 MiB" ], "fan": "unreadable on this board (fan.speed N/A); the operator's turbo fan per OPERATOR-CONDITIONS.md", "how": "nvidia-smi -i ; enforced.power.limit and clocks.max.sm READ, never typed; power.limit reads N/A on this board and the board refuses -pl", "loadavg": "2.16 2.51 2.65 2/1899 138694", "nvidia_powerd": "active", "stamp": "2026-09-23T21:26:41Z" }, "rapl_idle_resident": { "domains": { "core": { "joules": 84.195219, "mean_watts": 10.215453577169152, "samples": 9, "seconds": 8.24194622039795, "wraps": 0 }, "package-0": { "joules": 148.82097, "mean_watts": 18.05880390022705, "samples": 9, "seconds": 8.240909576416016, "wraps": 0 }, "psys": { "joules": 2.141413, "mean_watts": 0.2598071872127849, "samples": 9, "seconds": 8.242316246032715, "wraps": 0 }, "uncore": { "joules": 0.202941, "mean_watts": 0.024622920043686207, "samples": 9, "seconds": 8.241955041885376, "wraps": 0 } }, "instrument": "RAPL via /run/rapl/energy (root sampler, 1 Hz; R-10d): psys = the platform's own reading (not wall); package-0 = the CPU package" }, "rapl_task": { "domains": { "core": { "joules": 1715.491641, "mean_watts": 20.035969745223376, "samples": 84, "seconds": 85.62059450149536, "wraps": 0 }, "package-0": { "joules": 2428.107432, "mean_watts": 28.358511319095488, "samples": 84, "seconds": 85.62182283401489, "wraps": 0 }, "psys": { "joules": 22.623782, "mean_watts": 0.264235324225827, "samples": 84, "seconds": 85.61982417106628, "wraps": 0 }, "uncore": { "joules": 2.251764, "mean_watts": 0.026299579287708263, "samples": 84, "seconds": 85.6197726726532, "wraps": 0 } }, "instrument": "RAPL via /run/rapl/energy (root sampler, 1 Hz; R-10d): psys = the platform's own reading (not wall); package-0 = the CPU package" }, "seat_requests_task": { "gin_lines": 0, "how": "journalctl -u ollama (the system daemon :withheld-2026-10-14), read-only; a GIN line is one request it served", "lines": [], "since": "2026-09-23 21:26:51 UTC", "until": "2026-09-23 21:28:19 UTC" }, "summary": { "accuracy": 0.9074074074074074, "brier": 0.18518518518518517, "correct": 98, "errors": 0, "eval_tokens_total": 448, "floored_rows": 0, "letters_floored_total": 0, "median_seconds": 0.8071531855021021, "n": 108, "n_labelled": 108, "n_reached": 108, "p95_seconds": 0.9006965850130655, "prompt_tokens_total": 30905, "refused": 0, "refused_by_reason": {}, "reliability": [ { "accuracy": null, "correct": 0, "gap": null, "hi": 0.1, "lo": 0.0, "mean_p": null, "n": 0, "sum_p": 0.0 }, { "accuracy": null, "correct": 0, "gap": null, "hi": 0.2, "lo": 0.1, "mean_p": null, "n": 0, "sum_p": 0.0 }, { "accuracy": null, "correct": 0, "gap": null, "hi": 0.3, "lo": 0.2, "mean_p": null, "n": 0, "sum_p": 0.0 }, { "accuracy": null, "correct": 0, "gap": null, "hi": 0.4, "lo": 0.3, "mean_p": null, "n": 0, "sum_p": 0.0 }, { "accuracy": null, "correct": 0, "gap": null, "hi": 0.5, "lo": 0.4, "mean_p": null, "n": 0, "sum_p": 0.0 }, { "accuracy": null, "correct": 0, "gap": null, "hi": 0.6, "lo": 0.5, "mean_p": null, "n": 0, "sum_p": 0.0 }, { "accuracy": null, "correct": 0, "gap": null, "hi": 0.7, "lo": 0.6, "mean_p": null, "n": 0, "sum_p": 0.0 }, { "accuracy": null, "correct": 0, "gap": null, "hi": 0.8, "lo": 0.7, "mean_p": null, "n": 0, "sum_p": 0.0 }, { "accuracy": null, "correct": 0, "gap": null, "hi": 0.9, "lo": 0.8, "mean_p": null, "n": 0, "sum_p": 0.0 }, { "accuracy": 0.9074074074074074, "correct": 98, "gap": 0.09259259259259256, "hi": 1.0, "lo": 0.9, "mean_p": 1.0, "n": 108, "sum_p": 108.0 } ], "unparsed": 0 }, "wall_seconds": 87.47624816300231 }, "d": { "board_idle_window_epoch": [ 1790198901.171, 1790198910.176 ], "board_task_window_epoch": [ 1790198911.204, 1790198963.226 ], "energy": { "cards": [ "" ], "decisions": 48, "joules": 4379.630496305227, "joules_per_decision": 91.2423020063589, "mean_watts": 84.18804525090856, "net_joules_per_decision": 66.01581729759484, "peak_watts": 99.26, "per_card": { "": { "joules": 4379.630496305227, "mean_watts": 84.18804525090856, "min_watts": 23.21, "peak_watts": 99.26, "samples": 53, "seconds": 52.0220000743866 } }, "samples": 53, "seconds": 52.0220000743866 }, "idle": { "cards": [ "" ], "joules": 209.60162764787674, "mean_watts": 23.276138254762255, "peak_watts": 23.39, "per_card": { "": { "joules": 209.60162764787674, "mean_watts": 23.276138254762255, "min_watts": 23.14, "peak_watts": 23.39, "samples": 10, "seconds": 9.005000114440918 } }, "samples": 10, "seconds": 9.005000114440918, "what": "the draw of these cards immediately before the arm, with the model resident and no work in flight" }, "posture_end": { "ac_online": "1", "battery": { "capacity": "93", "status": "Not charging" }, "boards_on_bus": [ "0, , NVIDIA GeForce RTX 5090 Laptop GPU, withheld-2026-10-14" ], "card": { "clocks.max.sm": "3090", "clocks.mem": "14001", "clocks.sm": "1590", "clocks_throttle_reasons.active": "0x0000000000000000", "driver_version": "595.91.07", "enforced.power.limit": "150.00", "memory.total": "24463", "memory.used": "21121", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "pci.bus_id": "withheld-2026-10-14", "pcie.link.gen.current": "5", "pcie.link.gen.max": "5", "pcie.link.width.current": "8", "pcie.link.width.max": "16", "persistence_mode": "Enabled", "power.default_limit": "95.00", "power.draw": "45.92", "power.limit": "[N/A]", "power.max_limit": "175.00", "pstate": "P0", "temperature.gpu": "46", "utilization.gpu": "0", "uuid": "", "vbios_version": "98.03.A2.00.21" }, "compute_apps": [ ", 137540, /workshop/bench-coi-2026-09-23/ollama-0.32.13/lib/ollama/llama-server, 21098 MiB" ], "fan": "unreadable on this board (fan.speed N/A); the operator's turbo fan per OPERATOR-CONDITIONS.md", "how": "nvidia-smi -i ; enforced.power.limit and clocks.max.sm READ, never typed; power.limit reads N/A on this board and the board refuses -pl", "loadavg": "2.80 2.64 2.67 2/1902 160912", "nvidia_powerd": "active", "stamp": "2026-09-23T21:29:25Z" }, "posture_start": { "ac_online": "1", "battery": { "capacity": "93", "status": "Not charging" }, "boards_on_bus": [ "0, , NVIDIA GeForce RTX 5090 Laptop GPU, withheld-2026-10-14" ], "card": { "clocks.max.sm": "3090", "clocks.mem": "14001", "clocks.sm": "1590", "clocks_throttle_reasons.active": "0x0000000000000000", "driver_version": "595.91.07", "enforced.power.limit": "150.00", "memory.total": "24463", "memory.used": "21117", "name": "NVIDIA GeForce RTX 5090 Laptop GPU", "pci.bus_id": "withheld-2026-10-14", "pcie.link.gen.current": "5", "pcie.link.gen.max": "5", "pcie.link.width.current": "8", "pcie.link.width.max": "16", "persistence_mode": "Enabled", "power.default_limit": "95.00", "power.draw": "23.35", "power.limit": "[N/A]", "power.max_limit": "175.00", "pstate": "P0", "temperature.gpu": "44", "utilization.gpu": "0", "uuid": "", "vbios_version": "98.03.A2.00.21" }, "compute_apps": [ ", 137540, /workshop/bench-coi-2026-09-23/ollama-0.32.13/lib/ollama/llama-server, 21094 MiB" ], "fan": "unreadable on this board (fan.speed N/A); the operator's turbo fan per OPERATOR-CONDITIONS.md", "how": "nvidia-smi -i ; enforced.power.limit and clocks.max.sm READ, never typed; power.limit reads N/A on this board and the board refuses -pl", "loadavg": "2.91 2.64 2.68 5/1898 149393", "nvidia_powerd": "active", "stamp": "2026-09-23T21:28:21Z" }, "rapl_idle_resident": { "domains": { "core": { "joules": 65.202165, "mean_watts": 9.031674716295102, "samples": 8, "seconds": 7.219277381896973, "wraps": 0 }, "package-0": { "joules": 123.907154, "mean_watts": 17.164258911744597, "samples": 8, "seconds": 7.218904972076416, "wraps": 0 }, "psys": { "joules": 1.898005, "mean_watts": 0.26287545803765605, "samples": 8, "seconds": 7.220168113708496, "wraps": 0 }, "uncore": { "joules": 0.182495, "mean_watts": 0.025277051892852627, "samples": 8, "seconds": 7.219789743423462, "wraps": 0 } }, "instrument": "RAPL via /run/rapl/energy (root sampler, 1 Hz; R-10d): psys = the platform's own reading (not wall); package-0 = the CPU package" }, "rapl_task": { "domains": { "core": { "joules": 1283.496592, "mean_watts": 24.90457414886469, "samples": 51, "seconds": 51.53658056259155, "wraps": 0 }, "package-0": { "joules": 1732.18099, "mean_watts": 33.60992398663518, "samples": 51, "seconds": 51.5377836227417, "wraps": 0 }, "psys": { "joules": 13.765346, "mean_watts": 0.26710221738838036, "samples": 51, "seconds": 51.53587317466736, "wraps": 0 }, "uncore": { "joules": 1.434749, "mean_watts": 0.027839554789882953, "samples": 51, "seconds": 51.53634858131409, "wraps": 0 } }, "instrument": "RAPL via /run/rapl/energy (root sampler, 1 Hz; R-10d): psys = the platform's own reading (not wall); package-0 = the CPU package" }, "seat_requests_task": { "gin_lines": 0, "how": "journalctl -u ollama (the system daemon :withheld-2026-10-14), read-only; a GIN line is one request it served", "lines": [], "since": "2026-09-23 21:28:31 UTC", "until": "2026-09-23 21:29:24 UTC" }, "summary": { "accuracy": 0.8333333333333334, "brier": 0.3333333333333333, "correct": 40, "errors": 0, "eval_tokens_total": 146, "floored_rows": 0, "letters_floored_total": 0, "median_seconds": 1.0944841890013777, "n": 48, "n_labelled": 48, "n_reached": 48, "p95_seconds": 1.1680901319923578, "prompt_tokens_total": 24760, "refused": 0, "refused_by_reason": {}, "reliability": [ { "accuracy": null, "correct": 0, "gap": null, "hi": 0.1, "lo": 0.0, "mean_p": null, "n": 0, "sum_p": 0.0 }, { "accuracy": null, "correct": 0, "gap": null, "hi": 0.2, "lo": 0.1, "mean_p": null, "n": 0, "sum_p": 0.0 }, { "accuracy": null, "correct": 0, "gap": null, "hi": 0.3, "lo": 0.2, "mean_p": null, "n": 0, "sum_p": 0.0 }, { "accuracy": null, "correct": 0, "gap": null, "hi": 0.4, "lo": 0.3, "mean_p": null, "n": 0, "sum_p": 0.0 }, { "accuracy": null, "correct": 0, "gap": null, "hi": 0.5, "lo": 0.4, "mean_p": null, "n": 0, "sum_p": 0.0 }, { "accuracy": null, "correct": 0, "gap": null, "hi": 0.6, "lo": 0.5, "mean_p": null, "n": 0, "sum_p": 0.0 }, { "accuracy": null, "correct": 0, "gap": null, "hi": 0.7, "lo": 0.6, "mean_p": null, "n": 0, "sum_p": 0.0 }, { "accuracy": null, "correct": 0, "gap": null, "hi": 0.8, "lo": 0.7, "mean_p": null, "n": 0, "sum_p": 0.0 }, { "accuracy": null, "correct": 0, "gap": null, "hi": 0.9, "lo": 0.8, "mean_p": null, "n": 0, "sum_p": 0.0 }, { "accuracy": 0.8333333333333334, "correct": 40, "gap": 0.16666666666666663, "hi": 1.0, "lo": 0.9, "mean_p": 1.0, "n": 48, "sum_p": 48.0 } ], "unparsed": 0 }, "wall_seconds": 52.48190394400444 } }, "temp_guard": { "max_temp_c": 55.0, "rule": ">= 87 C on 6 consecutive 5 s reads", "tripped": null }, "think": "off", "think_budget": 8, "util_before_pct": [ 0.0, 0.0, 0.0 ], "warm": { "done_reason": "length", "load_ms": 4874.471602, "runner_pids": [ 137540, 1739347 ], "seconds": 4.989034255995648 } }