{"row_id": "R-LS-IQ3-24G/speed/a1/run-001", "arm": "R-LS-IQ3-24G", "engine": "llama.cpp", "engine_commit": "3cf03257f219afbe7334045ff7c6a06ac68c627d", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/workshop/card24-strata/build/llama.cpp-3cf03257-sm86/bin/llama-server", "-m", "/workshop/bench-flashnext/models/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf", "--host", "<loopback>", "--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": 22.276, "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": 22.275781159231652, "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": "0.00.296.741 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 64"}, "prompt_tok_s_server": 202.89561894695365, "prompt_tok_s_client": 202.8, "ttft_ms": 19914.273, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 19901.859, "prompt_per_token_ms": 4.9286426448737, "prompt_per_second": 202.89561894695365, "predicted_n": 256, "predicted_ms": 11447.41, "predicted_per_token_ms": 44.89180392156862, "predicted_per_second": 22.275781159231652}, "short_completion": false, "cache_flag": null, "completion_sha256": "53d2b3126f416897f9df8847a6aec5e09743db0febbdc6bb07e8cf5f66af7926", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1791538610.56, "t1": 1791538641.922, "seconds": 31.362, "host_samples": 0, "gpu_samples": 31}, "board": {"joules": 4943.13, "mean_w": 164.74, "max_w": 186.46, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": null, "label": "RAPL package-0: NOT WALL", "why": "fewer than 2 readable samples of /run/rapl/energy in the window"}, "data_reads_per_s": {"mean": null, "max": null}, "data_read_mb_per_s": {"mean": null}, "temps": {"gpu_max_c": 55.0, "ssd_max_c": null}, "vram": {"max_used_mib": 22396.0, "baseline_mib": 1.0, "growth_mib": 22395.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": null, "vmpin_max_bytes": null, "vmrss_max_bytes": null, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": null, "claude_procs_max": null}, "memory": {"mem_available_gib_min": null, "user_psi_full_avg10_max": null, "psi_full_avg10_max": null, "unit_psi_full_avg10_max": null, "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": null}, "power_limit": {"min_w": 300.0, "max_w": 300.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "R-LS-IQ3-24G/speed/a1/run-002", "arm": "R-LS-IQ3-24G", "engine": "llama.cpp", "engine_commit": "3cf03257f219afbe7334045ff7c6a06ac68c627d", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/workshop/card24-strata/build/llama.cpp-3cf03257-sm86/bin/llama-server", "-m", "/workshop/bench-flashnext/models/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf", "--host", "<loopback>", "--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": 22.49, "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": 22.490512972439685, "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": "0.00.296.741 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 64"}, "prompt_tok_s_server": 239.72701753765207, "prompt_tok_s_client": 239.4, "ttft_ms": 16866.494, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 16844.159, "prompt_per_token_ms": 4.171411342248637, "prompt_per_second": 239.72701753765207, "predicted_n": 256, "predicted_ms": 11338.114, "predicted_per_token_ms": 44.463192156862746, "predicted_per_second": 22.490512972439685}, "short_completion": false, "cache_flag": null, "completion_sha256": "b2fea38583808d43277db125a4e179cd6b26c8f5e941d23b7bd6f9988dbf6d20", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1791538641.928, "t1": 1791538670.133, "seconds": 28.205, "host_samples": 0, "gpu_samples": 29}, "board": {"joules": 4722.14, "mean_w": 168.62, "max_w": 188.85, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": null, "label": "RAPL package-0: NOT WALL", "why": "fewer than 2 readable samples of /run/rapl/energy in the window"}, "data_reads_per_s": {"mean": null, "max": null}, "data_read_mb_per_s": {"mean": null}, "temps": {"gpu_max_c": 53.0, "ssd_max_c": null}, "vram": {"max_used_mib": 22396.0, "baseline_mib": 1.0, "growth_mib": 22395.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": null, "vmpin_max_bytes": null, "vmrss_max_bytes": null, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": null, "claude_procs_max": null}, "memory": {"mem_available_gib_min": null, "user_psi_full_avg10_max": null, "psi_full_avg10_max": null, "unit_psi_full_avg10_max": null, "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": null}, "power_limit": {"min_w": 300.0, "max_w": 300.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "R-LS-IQ3-24G/speed/a1/run-003", "arm": "R-LS-IQ3-24G", "engine": "llama.cpp", "engine_commit": "3cf03257f219afbe7334045ff7c6a06ac68c627d", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/workshop/card24-strata/build/llama.cpp-3cf03257-sm86/bin/llama-server", "-m", "/workshop/bench-flashnext/models/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf", "--host", "<loopback>", "--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": 22.611, "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": 22.61092520891165, "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": "0.00.296.741 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 64"}, "prompt_tok_s_server": 240.06867462293965, "prompt_tok_s_client": 239.9, "ttft_ms": 16831.866, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 16820.187, "prompt_per_token_ms": 4.165474739970283, "prompt_per_second": 240.06867462293965, "predicted_n": 256, "predicted_ms": 11277.734, "predicted_per_token_ms": 44.22640784313726, "predicted_per_second": 22.61092520891165}, "short_completion": false, "cache_flag": null, "completion_sha256": "83316581364a0d7a32aadd157727670690c843988addb1ac33abda79fd85706b", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1791538670.139, "t1": 1791538698.248, "seconds": 28.11, "host_samples": 0, "gpu_samples": 28}, "board": {"joules": 4550.37, "mean_w": 168.5, "max_w": 187.85, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": null, "label": "RAPL package-0: NOT WALL", "why": "fewer than 2 readable samples of /run/rapl/energy in the window"}, "data_reads_per_s": {"mean": null, "max": null}, "data_read_mb_per_s": {"mean": null}, "temps": {"gpu_max_c": 54.0, "ssd_max_c": null}, "vram": {"max_used_mib": 22396.0, "baseline_mib": 1.0, "growth_mib": 22395.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": null, "vmpin_max_bytes": null, "vmrss_max_bytes": null, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": null, "claude_procs_max": null}, "memory": {"mem_available_gib_min": null, "user_psi_full_avg10_max": null, "psi_full_avg10_max": null, "unit_psi_full_avg10_max": null, "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": null}, "power_limit": {"min_w": 300.0, "max_w": 300.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "R-LS-IQ3-24G/speed/a1/run-004", "arm": "R-LS-IQ3-24G", "engine": "llama.cpp", "engine_commit": "3cf03257f219afbe7334045ff7c6a06ac68c627d", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/workshop/card24-strata/build/llama.cpp-3cf03257-sm86/bin/llama-server", "-m", "/workshop/bench-flashnext/models/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf", "--host", "<loopback>", "--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": 19.582, "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": 19.583445522156865, "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": "0.00.296.741 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 64"}, "prompt_tok_s_server": 216.04875977336908, "prompt_tok_s_client": 215.9, "ttft_ms": 148378.786, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 148290.599, "prompt_per_token_ms": 4.628584774330482, "prompt_per_second": 216.04875977336908, "predicted_n": 256, "predicted_ms": 13021.202, "predicted_per_token_ms": 51.06353725490196, "predicted_per_second": 19.583445522156865}, "short_completion": false, "cache_flag": null, "completion_sha256": "5d8180f3016cf0ceee27c984b9767a831a59cb80fa420fc2d570bf2b9398dcf5", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1791538698.254, "t1": 1791538859.655, "seconds": 161.401, "host_samples": 0, "gpu_samples": 161}, "board": {"joules": 28365.3, "mean_w": 177.26, "max_w": 206.44, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": null, "label": "RAPL package-0: NOT WALL", "why": "fewer than 2 readable samples of /run/rapl/energy in the window"}, "data_reads_per_s": {"mean": null, "max": null}, "data_read_mb_per_s": {"mean": null}, "temps": {"gpu_max_c": 54.0, "ssd_max_c": null}, "vram": {"max_used_mib": 22398.0, "baseline_mib": 1.0, "growth_mib": 22397.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": null, "vmpin_max_bytes": null, "vmrss_max_bytes": null, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": null, "claude_procs_max": null}, "memory": {"mem_available_gib_min": null, "user_psi_full_avg10_max": null, "psi_full_avg10_max": null, "unit_psi_full_avg10_max": null, "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": null}, "power_limit": {"min_w": 300.0, "max_w": 300.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "R-LS-IQ3-24G/speed/a1/run-005", "arm": "R-LS-IQ3-24G", "engine": "llama.cpp", "engine_commit": "3cf03257f219afbe7334045ff7c6a06ac68c627d", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/workshop/card24-strata/build/llama.cpp-3cf03257-sm86/bin/llama-server", "-m", "/workshop/bench-flashnext/models/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf", "--host", "<loopback>", "--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": 19.62, "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": 19.621820572532407, "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": "0.00.296.741 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 64"}, "prompt_tok_s_server": 232.87215251228878, "prompt_tok_s_client": 232.7, "ttft_ms": 137667.239, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 137577.635, "prompt_per_token_ms": 4.29420172919658, "prompt_per_second": 232.87215251228878, "predicted_n": 256, "predicted_ms": 12995.736, "predicted_per_token_ms": 50.963670588235296, "predicted_per_second": 19.621820572532407}, "short_completion": false, "cache_flag": null, "completion_sha256": "74992ed9477fe04610bf1202918b31fc05a2657c383dc2e933dde647a026bdaa", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1791538859.658, "t1": 1791539010.322, "seconds": 150.664, "host_samples": 0, "gpu_samples": 151}, "board": {"joules": 27337.31, "mean_w": 182.23, "max_w": 208.09, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": null, "label": "RAPL package-0: NOT WALL", "why": "fewer than 2 readable samples of /run/rapl/energy in the window"}, "data_reads_per_s": {"mean": null, "max": null}, "data_read_mb_per_s": {"mean": null}, "temps": {"gpu_max_c": 57.0, "ssd_max_c": null}, "vram": {"max_used_mib": 22398.0, "baseline_mib": 1.0, "growth_mib": 22397.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": null, "vmpin_max_bytes": null, "vmrss_max_bytes": null, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": null, "claude_procs_max": null}, "memory": {"mem_available_gib_min": null, "user_psi_full_avg10_max": null, "psi_full_avg10_max": null, "unit_psi_full_avg10_max": null, "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": null}, "power_limit": {"min_w": 300.0, "max_w": 300.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "R-LS-IQ3-24G/speed/a1/run-006", "arm": "R-LS-IQ3-24G", "engine": "llama.cpp", "engine_commit": "3cf03257f219afbe7334045ff7c6a06ac68c627d", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/workshop/card24-strata/build/llama.cpp-3cf03257-sm86/bin/llama-server", "-m", "/workshop/bench-flashnext/models/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf", "--host", "<loopback>", "--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": 19.686, "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": 19.687542487453825, "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": "0.00.296.741 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 64"}, "prompt_tok_s_server": 232.27690022322378, "prompt_tok_s_client": 232.1, "ttft_ms": 138019.73, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 137930.203, "prompt_per_token_ms": 4.305206411136775, "prompt_per_second": 232.27690022322378, "predicted_n": 256, "predicted_ms": 12952.353, "predicted_per_token_ms": 50.79354117647058, "predicted_per_second": 19.687542487453825}, "short_completion": false, "cache_flag": null, "completion_sha256": "c2a8c46aed307b3e41868991ce627535927881d8328f69f3b56def08f64427ba", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1791539010.324, "t1": 1791539161.298, "seconds": 150.974, "host_samples": 0, "gpu_samples": 151}, "board": {"joules": 27351.12, "mean_w": 182.32, "max_w": 208.99, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": null, "label": "RAPL package-0: NOT WALL", "why": "fewer than 2 readable samples of /run/rapl/energy in the window"}, "data_reads_per_s": {"mean": null, "max": null}, "data_read_mb_per_s": {"mean": null}, "temps": {"gpu_max_c": 56.0, "ssd_max_c": null}, "vram": {"max_used_mib": 22398.0, "baseline_mib": 1.0, "growth_mib": 22397.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": null, "vmpin_max_bytes": null, "vmrss_max_bytes": null, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": null, "claude_procs_max": null}, "memory": {"mem_available_gib_min": null, "user_psi_full_avg10_max": null, "psi_full_avg10_max": null, "unit_psi_full_avg10_max": null, "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": null}, "power_limit": {"min_w": 300.0, "max_w": 300.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "R-LS-IQ3-24G/speed/a1/run-007", "arm": "R-LS-IQ3-24G", "engine": "llama.cpp", "engine_commit": "3cf03257f219afbe7334045ff7c6a06ac68c627d", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/workshop/card24-strata/build/llama.cpp-3cf03257-sm86/bin/llama-server", "-m", "/workshop/bench-flashnext/models/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf", "--host", "<loopback>", "--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": 13.931, "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": 13.933692997466855, "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": "0.00.296.741 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 64"}, "prompt_tok_s_server": 211.54335816540018, "prompt_tok_s_client": 211.5, "ttft_ms": 605490.983, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 605251.808, "prompt_per_token_ms": 4.727163304357334, "prompt_per_second": 211.54335816540018, "predicted_n": 256, "predicted_ms": 18300.963, "predicted_per_token_ms": 71.76848235294118, "predicted_per_second": 13.933692997466855}, "short_completion": false, "cache_flag": null, "completion_sha256": "e2353fbe742c4222608638f97afda211bdd1e6256c8cc56d24f0ace38db034b5", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1791539161.31, "t1": 1791539785.106, "seconds": 623.796, "host_samples": 0, "gpu_samples": 623}, "board": {"joules": 112949.96, "mean_w": 181.57, "max_w": 263.83, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": null, "label": "RAPL package-0: NOT WALL", "why": "fewer than 2 readable samples of /run/rapl/energy in the window"}, "data_reads_per_s": {"mean": null, "max": null}, "data_read_mb_per_s": {"mean": null}, "temps": {"gpu_max_c": 57.0, "ssd_max_c": null}, "vram": {"max_used_mib": 22398.0, "baseline_mib": 1.0, "growth_mib": 22397.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": null, "vmpin_max_bytes": null, "vmrss_max_bytes": null, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": null, "claude_procs_max": null}, "memory": {"mem_available_gib_min": null, "user_psi_full_avg10_max": null, "psi_full_avg10_max": null, "unit_psi_full_avg10_max": null, "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": null}, "power_limit": {"min_w": 300.0, "max_w": 300.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "R-LS-IQ3-24G/speed/a1/run-008", "arm": "R-LS-IQ3-24G", "engine": "llama.cpp", "engine_commit": "3cf03257f219afbe7334045ff7c6a06ac68c627d", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/workshop/card24-strata/build/llama.cpp-3cf03257-sm86/bin/llama-server", "-m", "/workshop/bench-flashnext/models/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf", "--host", "<loopback>", "--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": 13.993, "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": 13.99539491135125, "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": "0.00.296.741 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 64"}, "prompt_tok_s_server": 226.47829956985873, "prompt_tok_s_client": 226.4, "ttft_ms": 565603.771, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 565338.932, "prompt_per_token_ms": 4.415434069839187, "prompt_per_second": 226.47829956985873, "predicted_n": 256, "predicted_ms": 18220.279, "predicted_per_token_ms": 71.45207450980392, "predicted_per_second": 13.99539491135125}, "short_completion": false, "cache_flag": null, "completion_sha256": "316991f5dc59c0d91b69e592113c707fa33fb64b7b51719b301f193d7a3f4e2f", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1791539785.112, "t1": 1791540368.94, "seconds": 583.828, "host_samples": 0, "gpu_samples": 584}, "board": {"joules": 108467.71, "mean_w": 186.03, "max_w": 265.46, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": null, "label": "RAPL package-0: NOT WALL", "why": "fewer than 2 readable samples of /run/rapl/energy in the window"}, "data_reads_per_s": {"mean": null, "max": null}, "data_read_mb_per_s": {"mean": null}, "temps": {"gpu_max_c": 57.0, "ssd_max_c": null}, "vram": {"max_used_mib": 22398.0, "baseline_mib": 1.0, "growth_mib": 22397.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": null, "vmpin_max_bytes": null, "vmrss_max_bytes": null, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": null, "claude_procs_max": null}, "memory": {"mem_available_gib_min": null, "user_psi_full_avg10_max": null, "psi_full_avg10_max": null, "unit_psi_full_avg10_max": null, "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": null}, "power_limit": {"min_w": 300.0, "max_w": 300.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}
{"row_id": "R-LS-IQ3-24G/speed/a1/run-009", "arm": "R-LS-IQ3-24G", "engine": "llama.cpp", "engine_commit": "3cf03257f219afbe7334045ff7c6a06ac68c627d", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/workshop/card24-strata/build/llama.cpp-3cf03257-sm86/bin/llama-server", "-m", "/workshop/bench-flashnext/models/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf", "--host", "<loopback>", "--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": 14.011, "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": 14.013752492181013, "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": "0.00.296.741 I cmn  common_param: system_info: n_threads = 8 (n_threads_batch = 8) / 64"}, "prompt_tok_s_server": 226.08212682537564, "prompt_tok_s_client": 226.0, "ttft_ms": 566567.801, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 566329.598, "prompt_per_token_ms": 4.423171411388895, "prompt_per_second": 226.08212682537564, "predicted_n": 256, "predicted_ms": 18196.411, "predicted_per_token_ms": 71.35847450980393, "predicted_per_second": 14.013752492181013}, "short_completion": false, "cache_flag": null, "completion_sha256": "ce7c3fb31977a6c0cdec812721d15e6cd578a944d237e968805865193eb1bca9", "completion_head": "withheld-2026-10-13", "strata": null, "telemetry": {"window": {"t0": 1791540368.944, "t1": 1791540953.712, "seconds": 584.768, "host_samples": 0, "gpu_samples": 585}, "board": {"joules": 108521.33, "mean_w": 185.8, "max_w": 266.21, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": null, "label": "RAPL package-0: NOT WALL", "why": "fewer than 2 readable samples of /run/rapl/energy in the window"}, "data_reads_per_s": {"mean": null, "max": null}, "data_read_mb_per_s": {"mean": null}, "temps": {"gpu_max_c": 59.0, "ssd_max_c": null}, "vram": {"max_used_mib": 22398.0, "baseline_mib": 1.0, "growth_mib": 22397.0, "rule": "max memory.used in the window minus the pre-load baseline (read before the unit started)"}, "pinned": {"vmlck_max_bytes": null, "vmpin_max_bytes": null, "vmrss_max_bytes": null, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": null, "claude_procs_max": null}, "memory": {"mem_available_gib_min": null, "user_psi_full_avg10_max": null, "psi_full_avg10_max": null, "unit_psi_full_avg10_max": null, "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": null}, "power_limit": {"min_w": 300.0, "max_w": 300.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}}