{"row_id": "L-IQ3-ub/speed/a1/run-001", "arm": "L-IQ3-ub", "engine": "llama.cpp", "engine_commit": "3cf03257f219afbe7334045ff7c6a06ac68c627d", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/build/llama.cpp-3cf03257/bin/llama-server", "-m", "/data/strata-bench/models/ISTA-ed59f920/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf", "--host", "127.0.0.1", "--port", "withheld-2026-10-13", "-ngl", "99", "--n-cpu-moe", "42", "-c", "131072", "-b", "4096", "-ub", "4096", "-fa", "on", "-ctk", "q8_0", "-ctv", "q8_0", "-lm", "mmap", "--lazy-mode", "on", "-fit", "off", "-np", "1", "--cache-ram", "0", "--no-webui", "--jinja", "-lv", "4", "-t", "8", "-tb", "8"], "ctx": "4K", "repeat": 0, "run": 1, "slice_sha256": "9c676dc1959897277fda62a2d77d01fc06c932a2e04641c753a24bde3e7d4a16", "prompt_sha256": "b42c6a181c3e02478e7e033585c5962f35eb046b680c0ae837ad8706a3c0674b", "nonce": "Ticket 85822155.", "requests_since_start": 1, "status": "ok", "fail_kind": null, "error": null, "decode_tok_s": 20.923, "decode_rule": "client clock: (completion_tokens - 1) / (t_finish - t_first_token) (the kit's one timing rule, seatlib.RULES['engine_timing']); WITHHELD (null, the computed value kept as decode_tok_s_withheld) when the token count fails its check", "decode_tok_s_server": 20.914879148497036, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "ok", "completion_tokens": 256, "token_frames": 257, "rule": "llama.cpp: token_frames == completion_tokens + 1 (one frame per token, plus the finish frame)"}, "token_frames": 257, "content_frames": 256, "frames": 259, "thinking_seen": false, "threads": {"n_threads": 8, "n_threads_batch": 8, "requested": {"t": 8, "tb": 8}, "line": "0.00.236.629 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 24"}, "prompt_tok_s_server": 222.05015198997467, "prompt_tok_s_client": 221.9, "ttft_ms": 18199.017, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 18185.081, "prompt_per_token_ms": 4.50348712233779, "prompt_per_second": 222.05015198997467, "predicted_n": 256, "predicted_ms": 12192.277, "predicted_per_token_ms": 47.812850980392156, "predicted_per_second": 20.914879148497036}, "short_completion": false, "cache_flag": null, "completion_sha256": "2d0af7deab99e2457b9901e027ca025db0b9ee87608cfb8a40292539bc27a3a2", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1791211456.747, "t1": 1791211487.133, "seconds": 30.387, "host_samples": 31, "gpu_samples": 31}, "board": {"joules": 1452.53, "mean_w": 48.36, "max_w": 76.23, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 1353.22, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 30.004}, "data_reads_per_s": {"mean": 7816.4, "max": 23982.1}, "data_read_mb_per_s": {"mean": 570.66}, "temps": {"gpu_max_c": 42.0, "ssd_max_c": 44.85}, "vram": {"max_used_mib": 21203.0, "baseline_mib": 15.0, "growth_mib": 21188.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": 0, "vmpin_max_bytes": 0, "vmrss_max_bytes": 47419691008, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 2.6, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 50.257, "user_psi_full_avg10_max": 8.88, "psi_full_avg10_max": 10.0, "unit_psi_full_avg10_max": 8.88, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 0}, "power_limit": {"min_w": 150.0, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "L-IQ3-ub/speed/a1/run-002", "arm": "L-IQ3-ub", "engine": "llama.cpp", "engine_commit": "3cf03257f219afbe7334045ff7c6a06ac68c627d", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/build/llama.cpp-3cf03257/bin/llama-server", "-m", "/data/strata-bench/models/ISTA-ed59f920/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf", "--host", "127.0.0.1", "--port", "withheld-2026-10-13", "-ngl", "99", "--n-cpu-moe", "42", "-c", "131072", "-b", "4096", "-ub", "4096", "-fa", "on", "-ctk", "q8_0", "-ctv", "q8_0", "-lm", "mmap", "--lazy-mode", "on", "-fit", "off", "-np", "1", "--cache-ram", "0", "--no-webui", "--jinja", "-lv", "4", "-t", "8", "-tb", "8"], "ctx": "4K", "repeat": 1, "run": 2, "slice_sha256": "9c676dc1959897277fda62a2d77d01fc06c932a2e04641c753a24bde3e7d4a16", "prompt_sha256": "9744fee0d5b25fee2aef36ad0bc7c5c4611f73d11f4e6bd9e5b84b17d93d319b", "nonce": "Ticket 42865306.", "requests_since_start": 2, "status": "ok", "fail_kind": null, "error": null, "decode_tok_s": 23.0, "decode_rule": "client clock: (completion_tokens - 1) / (t_finish - t_first_token) (the kit's one timing rule, seatlib.RULES['engine_timing']); WITHHELD (null, the computed value kept as decode_tok_s_withheld) when the token count fails its check", "decode_tok_s_server": 23.000642123808934, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "ok", "completion_tokens": 256, "token_frames": 257, "rule": "llama.cpp: token_frames == completion_tokens + 1 (one frame per token, plus the finish frame)"}, "token_frames": 257, "content_frames": 256, "frames": 259, "thinking_seen": false, "threads": {"n_threads": 8, "n_threads_batch": 8, "requested": {"t": 8, "tb": 8}, "line": "0.00.236.629 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 24"}, "prompt_tok_s_server": 533.7405870062586, "prompt_tok_s_client": 532.9, "ttft_ms": 7576.722, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 7565.473, "prompt_per_token_ms": 1.8735693412580485, "prompt_per_second": 533.7405870062586, "predicted_n": 256, "predicted_ms": 11086.647, "predicted_per_token_ms": 43.47704705882353, "predicted_per_second": 23.000642123808934}, "short_completion": false, "cache_flag": null, "completion_sha256": "bbb6c692948160168cd5e9bd31c355a0f39b22012e55f6408d47ef06212e9fae", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1791211487.135, "t1": 1791211505.799, "seconds": 18.664, "host_samples": 18, "gpu_samples": 18}, "board": {"joules": 1126.75, "mean_w": 66.2, "max_w": 90.37, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 877.59, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 17.002}, "data_reads_per_s": {"mean": 2879.6, "max": 25106.4}, "data_read_mb_per_s": {"mean": 24.85}, "temps": {"gpu_max_c": 43.0, "ssd_max_c": 42.85}, "vram": {"max_used_mib": 21303.0, "baseline_mib": 15.0, "growth_mib": 21288.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": 0, "vmpin_max_bytes": 0, "vmrss_max_bytes": 48332935168, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 2.95, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 49.842, "user_psi_full_avg10_max": 1.54, "psi_full_avg10_max": 1.72, "unit_psi_full_avg10_max": 1.2, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 0}, "power_limit": {"min_w": 136.53, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "L-IQ3-ub/speed/a1/run-003", "arm": "L-IQ3-ub", "engine": "llama.cpp", "engine_commit": "3cf03257f219afbe7334045ff7c6a06ac68c627d", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/build/llama.cpp-3cf03257/bin/llama-server", "-m", "/data/strata-bench/models/ISTA-ed59f920/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf", "--host", "127.0.0.1", "--port", "withheld-2026-10-13", "-ngl", "99", "--n-cpu-moe", "42", "-c", "131072", "-b", "4096", "-ub", "4096", "-fa", "on", "-ctk", "q8_0", "-ctv", "q8_0", "-lm", "mmap", "--lazy-mode", "on", "-fit", "off", "-np", "1", "--cache-ram", "0", "--no-webui", "--jinja", "-lv", "4", "-t", "8", "-tb", "8"], "ctx": "4K", "repeat": 2, "run": 3, "slice_sha256": "9c676dc1959897277fda62a2d77d01fc06c932a2e04641c753a24bde3e7d4a16", "prompt_sha256": "214ec62ff2b1550ea59fc2fa7cc264e79ceed8841f5fd91e1709a8b73be2fa8f", "nonce": "Ticket 90569497.", "requests_since_start": 3, "status": "ok", "fail_kind": null, "error": null, "decode_tok_s": 22.476, "decode_rule": "client clock: (completion_tokens - 1) / (t_finish - t_first_token) (the kit's one timing rule, seatlib.RULES['engine_timing']); WITHHELD (null, the computed value kept as decode_tok_s_withheld) when the token count fails its check", "decode_tok_s_server": 22.475279176019246, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "ok", "completion_tokens": 256, "token_frames": 257, "rule": "llama.cpp: token_frames == completion_tokens + 1 (one frame per token, plus the finish frame)"}, "token_frames": 257, "content_frames": 256, "frames": 259, "thinking_seen": false, "threads": {"n_threads": 8, "n_threads_batch": 8, "requested": {"t": 8, "tb": 8}, "line": "0.00.236.629 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 24"}, "prompt_tok_s_server": 799.8114362339538, "prompt_tok_s_client": 798.7, "ttft_ms": 5055.752, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 5048.69, "prompt_per_token_ms": 1.2502947003467062, "prompt_per_second": 799.8114362339538, "predicted_n": 256, "predicted_ms": 11345.799, "predicted_per_token_ms": 44.49332941176471, "predicted_per_second": 22.475279176019246}, "short_completion": false, "cache_flag": null, "completion_sha256": "ac026057b25631c13cd5c680188196474007eeb24f9860a0a1d5e3409e8e53b2", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1791211505.8, "t1": 1791211522.201, "seconds": 16.402, "host_samples": 17, "gpu_samples": 17}, "board": {"joules": 1176.48, "mean_w": 73.46, "max_w": 101.21, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 853.92, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 16.003}, "data_reads_per_s": {"mean": 153.6, "max": 794.8}, "data_read_mb_per_s": {"mean": 2.27}, "temps": {"gpu_max_c": 48.0, "ssd_max_c": 41.85}, "vram": {"max_used_mib": 21303.0, "baseline_mib": 15.0, "growth_mib": 21288.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": 0, "vmpin_max_bytes": 0, "vmrss_max_bytes": 48342134784, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 3.56, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 49.845, "user_psi_full_avg10_max": 0.25, "psi_full_avg10_max": 0.28, "unit_psi_full_avg10_max": 0.19, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 0}, "power_limit": {"min_w": 136.53, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "L-IQ3-ub/speed/a1/run-004", "arm": "L-IQ3-ub", "engine": "llama.cpp", "engine_commit": "3cf03257f219afbe7334045ff7c6a06ac68c627d", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/build/llama.cpp-3cf03257/bin/llama-server", "-m", "/data/strata-bench/models/ISTA-ed59f920/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf", "--host", "127.0.0.1", "--port", "withheld-2026-10-13", "-ngl", "99", "--n-cpu-moe", "42", "-c", "131072", "-b", "4096", "-ub", "4096", "-fa", "on", "-ctk", "q8_0", "-ctv", "q8_0", "-lm", "mmap", "--lazy-mode", "on", "-fit", "off", "-np", "1", "--cache-ram", "0", "--no-webui", "--jinja", "-lv", "4", "-t", "8", "-tb", "8"], "ctx": "32K", "repeat": 0, "run": 4, "slice_sha256": "b72b07858a845e76331ec38d46abae3c09d3e1cb6e4d968e887fa8b132c64977", "prompt_sha256": "507b7bc9961d968eaf07467339f314a3c76a036b2542852f07b9a638adcd5ad5", "nonce": "Ticket 10426037.", "requests_since_start": 4, "status": "ok", "fail_kind": null, "error": null, "decode_tok_s": 18.085, "decode_rule": "client clock: (completion_tokens - 1) / (t_finish - t_first_token) (the kit's one timing rule, seatlib.RULES['engine_timing']); WITHHELD (null, the computed value kept as decode_tok_s_withheld) when the token count fails its check", "decode_tok_s_server": 18.085892670369287, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "ok", "completion_tokens": 256, "token_frames": 257, "rule": "llama.cpp: token_frames == completion_tokens + 1 (one frame per token, plus the finish frame)"}, "token_frames": 257, "content_frames": 256, "frames": 259, "thinking_seen": false, "threads": {"n_threads": 8, "n_threads_batch": 8, "requested": {"t": 8, "tb": 8}, "line": "0.00.236.629 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 24"}, "prompt_tok_s_server": 640.9474002008708, "prompt_tok_s_client": 640.6, "ttft_ms": 50014.604, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 49985.381, "prompt_per_token_ms": 1.5601904301142393, "prompt_per_second": 640.9474002008708, "predicted_n": 256, "predicted_ms": 14099.387, "predicted_per_token_ms": 55.2917137254902, "predicted_per_second": 18.085892670369287}, "short_completion": false, "cache_flag": null, "completion_sha256": "d9ca4324a3684939a88cb9c18b7cdf2f5bdb91f052c1179c4ae133320c674e9e", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1791211522.202, "t1": 1791211586.317, "seconds": 64.115, "host_samples": 64, "gpu_samples": 64}, "board": {"joules": 5104.87, "mean_w": 81.01, "max_w": 100.81, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 2226.75, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 63.01}, "data_reads_per_s": {"mean": 2828.9, "max": 24601.0}, "data_read_mb_per_s": {"mean": 21.76}, "temps": {"gpu_max_c": 52.0, "ssd_max_c": 42.85}, "vram": {"max_used_mib": 21305.0, "baseline_mib": 15.0, "growth_mib": 21290.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": 0, "vmpin_max_bytes": 0, "vmrss_max_bytes": 50422419456, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 2.92, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 49.735, "user_psi_full_avg10_max": 0.04, "psi_full_avg10_max": 0.04, "unit_psi_full_avg10_max": 0.03, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 0}, "power_limit": {"min_w": 150.0, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "L-IQ3-ub/speed/a1/run-005", "arm": "L-IQ3-ub", "engine": "llama.cpp", "engine_commit": "3cf03257f219afbe7334045ff7c6a06ac68c627d", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/build/llama.cpp-3cf03257/bin/llama-server", "-m", "/data/strata-bench/models/ISTA-ed59f920/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf", "--host", "127.0.0.1", "--port", "withheld-2026-10-13", "-ngl", "99", "--n-cpu-moe", "42", "-c", "131072", "-b", "4096", "-ub", "4096", "-fa", "on", "-ctk", "q8_0", "-ctv", "q8_0", "-lm", "mmap", "--lazy-mode", "on", "-fit", "off", "-np", "1", "--cache-ram", "0", "--no-webui", "--jinja", "-lv", "4", "-t", "8", "-tb", "8"], "ctx": "32K", "repeat": 1, "run": 5, "slice_sha256": "b72b07858a845e76331ec38d46abae3c09d3e1cb6e4d968e887fa8b132c64977", "prompt_sha256": "c8a1ccb2149bdf9595fe8d97b7183fd9efd61c32259e907606ddbac7710955b2", "nonce": "Ticket 97542637.", "requests_since_start": 5, "status": "ok", "fail_kind": null, "error": null, "decode_tok_s": 19.875, "decode_rule": "client clock: (completion_tokens - 1) / (t_finish - t_first_token) (the kit's one timing rule, seatlib.RULES['engine_timing']); WITHHELD (null, the computed value kept as decode_tok_s_withheld) when the token count fails its check", "decode_tok_s_server": 19.87354347384396, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "ok", "completion_tokens": 256, "token_frames": 257, "rule": "llama.cpp: token_frames == completion_tokens + 1 (one frame per token, plus the finish frame)"}, "token_frames": 257, "content_frames": 256, "frames": 259, "thinking_seen": false, "threads": {"n_threads": 8, "n_threads_batch": 8, "requested": {"t": 8, "tb": 8}, "line": "0.00.236.629 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 24"}, "prompt_tok_s_server": 499.8772930591403, "prompt_tok_s_client": 499.6, "ttft_ms": 64123.627, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 64091.729, "prompt_per_token_ms": 2.000490948248954, "prompt_per_second": 499.8772930591403, "predicted_n": 256, "predicted_ms": 12831.129, "predicted_per_token_ms": 50.31815294117647, "predicted_per_second": 19.87354347384396}, "short_completion": false, "cache_flag": null, "completion_sha256": "eb24d1fc3419db7cb82fed57cd3189d839346166f39d22b10082217c45c64c9c", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1791211586.318, "t1": 1791211663.272, "seconds": 76.954, "host_samples": 77, "gpu_samples": 77}, "board": {"joules": 5408.24, "mean_w": 71.12, "max_w": 101.36, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 3825.19, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 76.01}, "data_reads_per_s": {"mean": 6155.6, "max": 40019.1}, "data_read_mb_per_s": {"mean": 74.12}, "temps": {"gpu_max_c": 54.0, "ssd_max_c": 43.85}, "vram": {"max_used_mib": 21305.0, "baseline_mib": 15.0, "growth_mib": 21290.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": 0, "vmpin_max_bytes": 0, "vmrss_max_bytes": 50422419456, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 2.72, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 47.601, "user_psi_full_avg10_max": 0.0, "psi_full_avg10_max": 0.54, "unit_psi_full_avg10_max": 0.0, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 0}, "power_limit": {"min_w": 135.15, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "L-IQ3-ub/speed/a1/run-006", "arm": "L-IQ3-ub", "engine": "llama.cpp", "engine_commit": "3cf03257f219afbe7334045ff7c6a06ac68c627d", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/build/llama.cpp-3cf03257/bin/llama-server", "-m", "/data/strata-bench/models/ISTA-ed59f920/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf", "--host", "127.0.0.1", "--port", "withheld-2026-10-13", "-ngl", "99", "--n-cpu-moe", "42", "-c", "131072", "-b", "4096", "-ub", "4096", "-fa", "on", "-ctk", "q8_0", "-ctv", "q8_0", "-lm", "mmap", "--lazy-mode", "on", "-fit", "off", "-np", "1", "--cache-ram", "0", "--no-webui", "--jinja", "-lv", "4", "-t", "8", "-tb", "8"], "ctx": "32K", "repeat": 2, "run": 6, "slice_sha256": "b72b07858a845e76331ec38d46abae3c09d3e1cb6e4d968e887fa8b132c64977", "prompt_sha256": "d6c964051c51baa9ef5cb009be0a921422c2b28350d2dbea610a6038552bdc18", "nonce": "Ticket 24502846.", "requests_since_start": 6, "status": "ok", "fail_kind": null, "error": null, "decode_tok_s": 19.393, "decode_rule": "client clock: (completion_tokens - 1) / (t_finish - t_first_token) (the kit's one timing rule, seatlib.RULES['engine_timing']); WITHHELD (null, the computed value kept as decode_tok_s_withheld) when the token count fails its check", "decode_tok_s_server": 19.393441594369055, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "ok", "completion_tokens": 256, "token_frames": 257, "rule": "llama.cpp: token_frames == completion_tokens + 1 (one frame per token, plus the finish frame)"}, "token_frames": 257, "content_frames": 256, "frames": 259, "thinking_seen": false, "threads": {"n_threads": 8, "n_threads_batch": 8, "requested": {"t": 8, "tb": 8}, "line": "0.00.236.629 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 24"}, "prompt_tok_s_server": 643.0411757100078, "prompt_tok_s_client": 642.6, "ttft_ms": 49860.234, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 49822.626, "prompt_per_token_ms": 1.5551103689368873, "prompt_per_second": 643.0411757100078, "predicted_n": 256, "predicted_ms": 13148.775, "predicted_per_token_ms": 51.56382352941176, "predicted_per_second": 19.393441594369055}, "short_completion": false, "cache_flag": null, "completion_sha256": "8e4e4e8a609a2cde76b2bf2dfe33e95757b9bf5d9978a890cf6b83a39fb4c6be", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1791211663.275, "t1": 1791211726.285, "seconds": 63.009, "host_samples": 63, "gpu_samples": 63}, "board": {"joules": 5116.05, "mean_w": 82.41, "max_w": 110.94, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 2528.79, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 62.008}, "data_reads_per_s": {"mean": 2844.9, "max": 23658.8}, "data_read_mb_per_s": {"mean": 15.33}, "temps": {"gpu_max_c": 57.0, "ssd_max_c": 44.85}, "vram": {"max_used_mib": 21305.0, "baseline_mib": 15.0, "growth_mib": 21290.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": 0, "vmpin_max_bytes": 0, "vmrss_max_bytes": 48496410624, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 3.6, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 48.594, "user_psi_full_avg10_max": 4.26, "psi_full_avg10_max": 4.16, "unit_psi_full_avg10_max": 0.72, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 0}, "power_limit": {"min_w": 135.0, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "L-IQ3-ub/speed/a1/run-007", "arm": "L-IQ3-ub", "engine": "llama.cpp", "engine_commit": "3cf03257f219afbe7334045ff7c6a06ac68c627d", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/build/llama.cpp-3cf03257/bin/llama-server", "-m", "/data/strata-bench/models/ISTA-ed59f920/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf", "--host", "127.0.0.1", "--port", "withheld-2026-10-13", "-ngl", "99", "--n-cpu-moe", "42", "-c", "131072", "-b", "4096", "-ub", "4096", "-fa", "on", "-ctk", "q8_0", "-ctv", "q8_0", "-lm", "mmap", "--lazy-mode", "on", "-fit", "off", "-np", "1", "--cache-ram", "0", "--no-webui", "--jinja", "-lv", "4", "-t", "8", "-tb", "8"], "ctx": "128K", "repeat": 0, "run": 7, "slice_sha256": "286827657f5df11459f4ee550246f7e0e06ce7af3238fe74a53f0aa4c12676ac", "prompt_sha256": "2491e963527b8a39f7fff1aa2d04dab5eb4dd3ca12c7417c451f6216c665972c", "nonce": "Ticket 14239051.", "requests_since_start": 7, "status": "ok", "fail_kind": null, "error": null, "decode_tok_s": 13.003, "decode_rule": "client clock: (completion_tokens - 1) / (t_finish - t_first_token) (the kit's one timing rule, seatlib.RULES['engine_timing']); WITHHELD (null, the computed value kept as decode_tok_s_withheld) when the token count fails its check", "decode_tok_s_server": 13.004069610835623, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "ok", "completion_tokens": 256, "token_frames": 257, "rule": "llama.cpp: token_frames == completion_tokens + 1 (one frame per token, plus the finish frame)"}, "token_frames": 257, "content_frames": 256, "frames": 259, "thinking_seen": false, "threads": {"n_threads": 8, "n_threads_batch": 8, "requested": {"t": 8, "tb": 8}, "line": "0.00.236.629 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 24"}, "prompt_tok_s_server": 626.7390593021529, "prompt_tok_s_client": 626.4, "ttft_ms": 204395.922, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 204290.762, "prompt_per_token_ms": 1.595560361458016, "prompt_per_second": 626.7390593021529, "predicted_n": 256, "predicted_ms": 19609.246, "predicted_per_token_ms": 76.89900392156862, "predicted_per_second": 13.004069610835623}, "short_completion": false, "cache_flag": null, "completion_sha256": "0051e34541393c1b29e6039b246f10b7f6c1c2dfe20985d133b22da5c484e980", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1791211726.288, "t1": 1791211950.295, "seconds": 224.007, "host_samples": 224, "gpu_samples": 223}, "board": {"joules": 19290.33, "mean_w": 86.8, "max_w": 112.65, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 8207.23, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 223.033}, "data_reads_per_s": {"mean": 2854.2, "max": 24619.8}, "data_read_mb_per_s": {"mean": 13.82}, "temps": {"gpu_max_c": 57.0, "ssd_max_c": 46.85}, "vram": {"max_used_mib": 21305.0, "baseline_mib": 15.0, "growth_mib": 21290.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": 0, "vmpin_max_bytes": 0, "vmrss_max_bytes": 50028134400, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 2.18, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 48.036, "user_psi_full_avg10_max": 0.02, "psi_full_avg10_max": 0.02, "unit_psi_full_avg10_max": 0.32, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 0}, "power_limit": {"min_w": 138.99, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "L-IQ3-ub/speed/a1/run-008", "arm": "L-IQ3-ub", "engine": "llama.cpp", "engine_commit": "3cf03257f219afbe7334045ff7c6a06ac68c627d", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/build/llama.cpp-3cf03257/bin/llama-server", "-m", "/data/strata-bench/models/ISTA-ed59f920/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf", "--host", "127.0.0.1", "--port", "withheld-2026-10-13", "-ngl", "99", "--n-cpu-moe", "42", "-c", "131072", "-b", "4096", "-ub", "4096", "-fa", "on", "-ctk", "q8_0", "-ctv", "q8_0", "-lm", "mmap", "--lazy-mode", "on", "-fit", "off", "-np", "1", "--cache-ram", "0", "--no-webui", "--jinja", "-lv", "4", "-t", "8", "-tb", "8"], "ctx": "128K", "repeat": 1, "run": 8, "slice_sha256": "286827657f5df11459f4ee550246f7e0e06ce7af3238fe74a53f0aa4c12676ac", "prompt_sha256": "94613594c1c300728a39aa8436592de0bdf88c05edeceb88cd9c6794de1fa097", "nonce": "Ticket 81078402.", "requests_since_start": 8, "status": "ok", "fail_kind": null, "error": null, "decode_tok_s": 13.044, "decode_rule": "client clock: (completion_tokens - 1) / (t_finish - t_first_token) (the kit's one timing rule, seatlib.RULES['engine_timing']); WITHHELD (null, the computed value kept as decode_tok_s_withheld) when the token count fails its check", "decode_tok_s_server": 13.045203163691976, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "ok", "completion_tokens": 256, "token_frames": 257, "rule": "llama.cpp: token_frames == completion_tokens + 1 (one frame per token, plus the finish frame)"}, "token_frames": 257, "content_frames": 256, "frames": 259, "thinking_seen": false, "threads": {"n_threads": 8, "n_threads_batch": 8, "requested": {"t": 8, "tb": 8}, "line": "0.00.236.629 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 24"}, "prompt_tok_s_server": 691.0919640038875, "prompt_tok_s_client": 690.7, "ttft_ms": 185370.583, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 185267.673, "prompt_per_token_ms": 1.4469854260877715, "prompt_per_second": 691.0919640038875, "predicted_n": 256, "predicted_ms": 19547.415, "predicted_per_token_ms": 76.65652941176471, "predicted_per_second": 13.045203163691976}, "short_completion": false, "cache_flag": null, "completion_sha256": "60449170e6875411a9ed5d897bd66ffd19d42849b4d37429fc46d74e1463d082", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1791211950.298, "t1": 1791212155.218, "seconds": 204.92, "host_samples": 205, "gpu_samples": 205}, "board": {"joules": 18912.05, "mean_w": 92.61, "max_w": 116.2, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 6463.33, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 204.032}, "data_reads_per_s": {"mean": 1475.6, "max": 24369.7}, "data_read_mb_per_s": {"mean": 6.07}, "temps": {"gpu_max_c": 57.0, "ssd_max_c": 46.85}, "vram": {"max_used_mib": 21305.0, "baseline_mib": 15.0, "growth_mib": 21290.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": 0, "vmpin_max_bytes": 0, "vmrss_max_bytes": 50180542464, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 1.33, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 49.092, "user_psi_full_avg10_max": 0.0, "psi_full_avg10_max": 0.18, "unit_psi_full_avg10_max": 0.0, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 0}, "power_limit": {"min_w": 150.0, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "L-IQ3-ub/speed/a1/run-009", "arm": "L-IQ3-ub", "engine": "llama.cpp", "engine_commit": "3cf03257f219afbe7334045ff7c6a06ac68c627d", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/build/llama.cpp-3cf03257/bin/llama-server", "-m", "/data/strata-bench/models/ISTA-ed59f920/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf", "--host", "127.0.0.1", "--port", "withheld-2026-10-13", "-ngl", "99", "--n-cpu-moe", "42", "-c", "131072", "-b", "4096", "-ub", "4096", "-fa", "on", "-ctk", "q8_0", "-ctv", "q8_0", "-lm", "mmap", "--lazy-mode", "on", "-fit", "off", "-np", "1", "--cache-ram", "0", "--no-webui", "--jinja", "-lv", "4", "-t", "8", "-tb", "8"], "ctx": "128K", "repeat": 2, "run": 9, "slice_sha256": "286827657f5df11459f4ee550246f7e0e06ce7af3238fe74a53f0aa4c12676ac", "prompt_sha256": "38d3c9a3177694d5b9006f1c813fddc9e6c5745048a59c545c41d199f780b364", "nonce": "Ticket 81461625.", "requests_since_start": 9, "status": "ok", "fail_kind": null, "error": null, "decode_tok_s": 13.482, "decode_rule": "client clock: (completion_tokens - 1) / (t_finish - t_first_token) (the kit's one timing rule, seatlib.RULES['engine_timing']); WITHHELD (null, the computed value kept as decode_tok_s_withheld) when the token count fails its check", "decode_tok_s_server": 13.483660817084465, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "ok", "completion_tokens": 256, "token_frames": 257, "rule": "llama.cpp: token_frames == completion_tokens + 1 (one frame per token, plus the finish frame)"}, "token_frames": 257, "content_frames": 256, "frames": 259, "thinking_seen": false, "threads": {"n_threads": 8, "n_threads_batch": 8, "requested": {"t": 8, "tb": 8}, "line": "0.00.236.629 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 24"}, "prompt_tok_s_server": 609.6924659430994, "prompt_tok_s_client": 609.4, "ttft_ms": 210102.832, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 210002.595, "prompt_per_token_ms": 1.64017116146114, "prompt_per_second": 609.6924659430994, "predicted_n": 256, "predicted_ms": 18911.778, "predicted_per_token_ms": 74.16383529411765, "predicted_per_second": 13.483660817084465}, "short_completion": false, "cache_flag": null, "completion_sha256": "b8af56e6b75cd7774171340925a4b78fd0be75976fc5d103b1e6c47cc35835f7", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1791212155.219, "t1": 1791212384.236, "seconds": 229.016, "host_samples": 229, "gpu_samples": 229}, "board": {"joules": 19466.61, "mean_w": 85.28, "max_w": 112.57, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 9630.92, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 228.037}, "data_reads_per_s": {"mean": 3236.6, "max": 24779.0}, "data_read_mb_per_s": {"mean": 13.26}, "temps": {"gpu_max_c": 58.0, "ssd_max_c": 46.85}, "vram": {"max_used_mib": 21305.0, "baseline_mib": 15.0, "growth_mib": 21290.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": 0, "vmpin_max_bytes": 0, "vmrss_max_bytes": 49823920128, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 2.28, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 47.65, "user_psi_full_avg10_max": 0.0, "psi_full_avg10_max": 0.0, "unit_psi_full_avg10_max": 0.0, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 0}, "power_limit": {"min_w": 150.0, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}