{"row_id": "L-IQ3-t8/speed/a1/run-001", "arm": "L-IQ3-t8", "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", "34", "-c", "131072", "-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": 24.424, "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": 24.418209817729075, "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": "0.00.224.969 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 24"}, "prompt_tok_s_server": 109.88478398074693, "prompt_tok_s_client": 109.9, "ttft_ms": 36757.588, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 36747.581, "prompt_per_token_ms": 9.100441059930658, "prompt_per_second": 109.88478398074693, "predicted_n": 256, "predicted_ms": 10443.026, "predicted_per_token_ms": 40.9530431372549, "predicted_per_second": 24.418209817729075}, "short_completion": false, "cache_flag": null, "completion_sha256": "fc2ce3cc51d1193c284c1b7d946fbd06bba246c971756f4f83f4cb47309acc83", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1790982218.036, "t1": 1790982265.235, "seconds": 47.198, "host_samples": 47, "gpu_samples": 47}, "board": {"joules": 2059.51, "mean_w": 44.75, "max_w": 76.97, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 1583.1, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 46.006}, "data_reads_per_s": {"mean": 5202.8, "max": 41606.0}, "data_read_mb_per_s": {"mean": 389.61}, "temps": {"gpu_max_c": 47.0, "ssd_max_c": 47.85}, "vram": {"max_used_mib": 22467.0, "baseline_mib": 15.0, "growth_mib": 22452.0, "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": 36361129984, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 1.32, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 50.977, "user_psi_full_avg10_max": 1.9, "psi_full_avg10_max": 4.37, "unit_psi_full_avg10_max": 5.92, "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": 143.0, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "L-IQ3-t8/speed/a1/run-002", "arm": "L-IQ3-t8", "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", "34", "-c", "131072", "-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": 23.96, "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": 23.960371237293838, "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": "0.00.224.969 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 24"}, "prompt_tok_s_server": 211.6818389458727, "prompt_tok_s_client": 211.6, "ttft_ms": 19082.624, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 19075.798, "prompt_per_token_ms": 4.72407082714215, "prompt_per_second": 211.6818389458727, "predicted_n": 256, "predicted_ms": 10642.573, "predicted_per_token_ms": 41.73558039215686, "predicted_per_second": 23.960371237293838}, "short_completion": false, "cache_flag": null, "completion_sha256": "7b824d095cc81409609a68f1f23002bc7a38516b6033e8c065c9619713692a8e", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1790982265.236, "t1": 1790982294.961, "seconds": 29.725, "host_samples": 30, "gpu_samples": 30}, "board": {"joules": 1774.22, "mean_w": 61.11, "max_w": 73.67, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 1333.84, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 29.003}, "data_reads_per_s": {"mean": 1687.1, "max": 7219.4}, "data_read_mb_per_s": {"mean": 11.25}, "temps": {"gpu_max_c": 48.0, "ssd_max_c": 46.85}, "vram": {"max_used_mib": 22467.0, "baseline_mib": 15.0, "growth_mib": 22452.0, "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": 36149415936, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 2.81, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 50.042, "user_psi_full_avg10_max": 4.64, "psi_full_avg10_max": 5.09, "unit_psi_full_avg10_max": 0.1, "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": 143.0, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "L-IQ3-t8/speed/a1/run-003", "arm": "L-IQ3-t8", "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", "34", "-c", "131072", "-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": 24.743, "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": 24.743206752586076, "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": "0.00.224.969 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 24"}, "prompt_tok_s_server": 236.00894496111258, "prompt_tok_s_client": 235.9, "ttft_ms": 17116.701, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 17109.521, "prompt_per_token_ms": 4.237127538385339, "prompt_per_second": 236.00894496111258, "predicted_n": 256, "predicted_ms": 10305.859, "predicted_per_token_ms": 40.41513333333334, "predicted_per_second": 24.743206752586076}, "short_completion": false, "cache_flag": null, "completion_sha256": "2f42cb414309477fc9d218ba3a726ae2762280c21580e79f45f37104fed104e0", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1790982294.963, "t1": 1790982322.386, "seconds": 27.423, "host_samples": 24, "gpu_samples": 27}, "board": {"joules": 1681.62, "mean_w": 64.62, "max_w": 78.8, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 1313.92, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 26.908}, "data_reads_per_s": {"mean": 1355.0, "max": 7395.8}, "data_read_mb_per_s": {"mean": 8.23}, "temps": {"gpu_max_c": 53.0, "ssd_max_c": 46.85}, "vram": {"max_used_mib": 22467.0, "baseline_mib": 15.0, "growth_mib": 22452.0, "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": 36312866816, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 3.78, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 50.132, "user_psi_full_avg10_max": 0.93, "psi_full_avg10_max": 1.02, "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": 141.15, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "L-IQ3-t8/speed/a1/run-004", "arm": "L-IQ3-t8", "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", "34", "-c", "131072", "-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": 20.938, "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": 20.938534071059472, "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": "0.00.224.969 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 24"}, "prompt_tok_s_server": 234.5073091685792, "prompt_tok_s_client": 234.5, "ttft_ms": 136648.21, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 136618.343, "prompt_per_token_ms": 4.264259410699794, "prompt_per_second": 234.5073091685792, "predicted_n": 256, "predicted_ms": 12178.503, "predicted_per_token_ms": 47.75883529411765, "predicted_per_second": 20.938534071059472}, "short_completion": false, "cache_flag": null, "completion_sha256": "e3cdde9fd2287da6e2ead8c329b64adbfc0655de9cf614bf07aec53e25b9827e", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1790982322.386, "t1": 1790982471.214, "seconds": 148.827, "host_samples": 148, "gpu_samples": 149}, "board": {"joules": 9681.01, "mean_w": 65.33, "max_w": 81.48, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 5113.47, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 147.024}, "data_reads_per_s": {"mean": 1318.1, "max": 7091.8}, "data_read_mb_per_s": {"mean": 16.63}, "temps": {"gpu_max_c": 54.0, "ssd_max_c": 48.85}, "vram": {"max_used_mib": 22469.0, "baseline_mib": 15.0, "growth_mib": 22454.0, "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": 39260258304, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 3.49, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 52.119, "user_psi_full_avg10_max": 0.05, "psi_full_avg10_max": 0.16, "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": 143.06, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "L-IQ3-t8/speed/a1/run-005", "arm": "L-IQ3-t8", "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", "34", "-c", "131072", "-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": 20.663, "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": 20.663293336760468, "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": "0.00.224.969 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 24"}, "prompt_tok_s_server": 258.6007616300139, "prompt_tok_s_client": 258.5, "ttft_ms": 123919.533, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 123889.813, "prompt_per_token_ms": 3.8669646357450524, "prompt_per_second": 258.6007616300139, "predicted_n": 256, "predicted_ms": 12340.724, "predicted_per_token_ms": 48.394996078431376, "predicted_per_second": 20.663293336760468}, "short_completion": false, "cache_flag": null, "completion_sha256": "51a9ac7d1a2ec33a49b748e3d7c3ebcaabf35c8e1bb868b4bbed0bdc8e231fe6", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1790982471.215, "t1": 1790982607.476, "seconds": 136.261, "host_samples": 137, "gpu_samples": 136}, "board": {"joules": 9370.89, "mean_w": 69.34, "max_w": 84.07, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 4947.69, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 136.022}, "data_reads_per_s": {"mean": 14.7, "max": 218.0}, "data_read_mb_per_s": {"mean": 0.21}, "temps": {"gpu_max_c": 56.0, "ssd_max_c": 48.85}, "vram": {"max_used_mib": 22469.0, "baseline_mib": 15.0, "growth_mib": 22454.0, "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": 39298437120, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 1.5, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 51.918, "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": 148.41, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "L-IQ3-t8/speed/a1/run-006", "arm": "L-IQ3-t8", "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", "34", "-c", "131072", "-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": 20.477, "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": 20.47805908474158, "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": "0.00.224.969 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 24"}, "prompt_tok_s_server": 255.14733954505104, "prompt_tok_s_client": 255.1, "ttft_ms": 125598.069, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 125566.663, "prompt_per_token_ms": 3.9193040451963292, "prompt_per_second": 255.14733954505104, "predicted_n": 256, "predicted_ms": 12452.352, "predicted_per_token_ms": 48.83275294117647, "predicted_per_second": 20.47805908474158}, "short_completion": false, "cache_flag": null, "completion_sha256": "eb1608506a384aeaf4265b814ad75ce38472e3731d5ed4905066b5ceb2dade56", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1790982607.477, "t1": 1790982745.528, "seconds": 138.051, "host_samples": 138, "gpu_samples": 138}, "board": {"joules": 9425.33, "mean_w": 68.77, "max_w": 83.29, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 5181.71, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 137.021}, "data_reads_per_s": {"mean": 10.8, "max": 205.9}, "data_read_mb_per_s": {"mean": 0.12}, "temps": {"gpu_max_c": 57.0, "ssd_max_c": 47.85}, "vram": {"max_used_mib": 22469.0, "baseline_mib": 15.0, "growth_mib": 22454.0, "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": 39322042368, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 1.56, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 51.969, "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"}}}
{"row_id": "L-IQ3-t8/speed/a1/run-007", "arm": "L-IQ3-t8", "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", "34", "-c", "131072", "-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": 14.33, "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": 14.328950790955274, "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": "0.00.224.969 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 24"}, "prompt_tok_s_server": 224.78334154474206, "prompt_tok_s_client": 224.7, "ttft_ms": 569707.45, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 569601.818, "prompt_per_token_ms": 4.448728242617368, "prompt_per_second": 224.78334154474206, "predicted_n": 256, "predicted_ms": 17796.139, "predicted_per_token_ms": 69.78878039215687, "predicted_per_second": 14.328950790955274}, "short_completion": false, "cache_flag": null, "completion_sha256": "13ad5f9309c4c83d46f057dca4743a2a50f1dd4f0a90c204407c75d0e2734413", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1790982745.531, "t1": 1790983333.034, "seconds": 587.503, "host_samples": 587, "gpu_samples": 587}, "board": {"joules": 38892.04, "mean_w": 66.29, "max_w": 107.81, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 21160.53, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 586.089}, "data_reads_per_s": {"mean": 1471.4, "max": 24183.1}, "data_read_mb_per_s": {"mean": 12.66}, "temps": {"gpu_max_c": 57.0, "ssd_max_c": 48.85}, "vram": {"max_used_mib": 22469.0, "baseline_mib": 15.0, "growth_mib": 22454.0, "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": 40517881856, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 2.04, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 50.932, "user_psi_full_avg10_max": 0.0, "psi_full_avg10_max": 1.48, "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-t8/speed/a1/run-008", "arm": "L-IQ3-t8", "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", "34", "-c", "131072", "-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": 14.194, "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": 14.195441615456845, "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": "0.00.224.969 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 24"}, "prompt_tok_s_server": 237.60825938980892, "prompt_tok_s_client": 237.6, "ttft_ms": 538969.293, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 538857.531, "prompt_per_token_ms": 4.208607910213454, "prompt_per_second": 237.60825938980892, "predicted_n": 256, "predicted_ms": 17963.513, "predicted_per_token_ms": 70.44514901960784, "predicted_per_second": 14.195441615456845}, "short_completion": false, "cache_flag": null, "completion_sha256": "d2b8e46b8829106f42ef5f8083aa697dc522ab62df36974c8374dfcc99104f75", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1790983333.038, "t1": 1790983889.973, "seconds": 556.935, "host_samples": 557, "gpu_samples": 556}, "board": {"joules": 37941.46, "mean_w": 68.28, "max_w": 104.67, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 19457.55, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 556.085}, "data_reads_per_s": {"mean": 632.0, "max": 6154.4}, "data_read_mb_per_s": {"mean": 2.62}, "temps": {"gpu_max_c": 57.0, "ssd_max_c": 48.85}, "vram": {"max_used_mib": 22469.0, "baseline_mib": 15.0, "growth_mib": 22454.0, "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": 41305264128, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 1.78, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 51.934, "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"}}}
{"row_id": "L-IQ3-t8/speed/a1/run-009", "arm": "L-IQ3-t8", "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", "34", "-c", "131072", "-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": 14.24, "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": 14.241562251348606, "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": "0.00.224.969 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 24"}, "prompt_tok_s_server": 234.92579051756965, "prompt_tok_s_client": 234.9, "ttft_ms": 545112.906, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 545010.404, "prompt_per_token_ms": 4.256663339503425, "prompt_per_second": 234.92579051756965, "predicted_n": 256, "predicted_ms": 17905.339, "predicted_per_token_ms": 70.21701568627451, "predicted_per_second": 14.241562251348606}, "short_completion": false, "cache_flag": null, "completion_sha256": "333e42b05df1010a52608c7670f9c4280b75b12127d2a707fd584f673930e472", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1790983889.974, "t1": 1790984452.994, "seconds": 563.02, "host_samples": 563, "gpu_samples": 563}, "board": {"joules": 38256.47, "mean_w": 67.99, "max_w": 111.05, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 19817.76, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 562.085}, "data_reads_per_s": {"mean": 845.3, "max": 6274.0}, "data_read_mb_per_s": {"mean": 3.57}, "temps": {"gpu_max_c": 57.0, "ssd_max_c": 47.85}, "vram": {"max_used_mib": 22469.0, "baseline_mib": 15.0, "growth_mib": 22454.0, "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": 41314074624, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 1.85, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 51.97, "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"}}}