{"row_id": "R-S139-IQ3-24G/speed/a1/run-001", "arm": "R-S139-IQ3-24G", "engine": "strata", "engine_commit": "6f32ec070f23ced9f50e704d854d775da52591ab", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/usr/bin/python3", "/workshop/bench-flashnext/strata-6f32ec07/serve/server.py", "--engine", "strata", "--config", "/workshop/card24-strata/results/R-S139-IQ3-24G/R-S139-IQ3-24G.run/strata-config.json", "--port", "withheld-2026-10-13"], "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": 88.015, "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": 87.4, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "agree at run time (text recount at summary)", "completion_tokens": 256, "against": {"engine_generated (the 'N generated' log line)": 256, "metrics output_tokens": 256}, "gaps": {"engine_generated (the 'N generated' log line)": 0, "metrics output_tokens": 0}, "unread": [], "tolerance": "3% of the reported count", "rule": "Strata: usage completion_tokens against the engine's 'N generated' line and /metrics output_tokens (a gap above 3 % is COUNT-MISMATCH), then against the kept text recounted on llama.cpp's tokenizer at summary time (the same 3 %)"}, "token_frames": 249, "content_frames": 248, "frames": 250, "thinking_seen": false, "threads": {"workers": 31, "host_thread": true, "line": "strata generate: 31 expert-pool workers + the host thread"}, "prompt_tok_s_server": 1872.9, "prompt_tok_s_client": 1826.5, "ttft_ms": 2210.818, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 2156.0, "prompt_per_token_ms": 0.534, "prompt_per_second": 1872.9, "predicted_n": 256, "predicted_ms": 2927.5, "predicted_per_token_ms": 11.436, "predicted_per_second": 87.4, "draft_n": 220, "draft_n_accepted": 163}, "short_completion": false, "cache_flag": null, "completion_sha256": "ec3a92b56c07d218872678857c1bf4ed9cd7e79691e7d736c84cec1362589a89", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.8740000000000001, "hits": 120532, "lookups": 137987, "hit_rate_line": "present", "offloaded": 12253, "pcie_share_log": 0.08199999999999999, "routed": 150240, "drafts_accepted": 163, "drafts_offered": 220, "engine_generated": 256, "engine_decode_tok_s": 87.4, "engine_prompt_tok_s": 1872.9, "reused": 0, "metrics": {"hit_rate": 0.874, "pcie_share": 0.082, "decode_tok_s": 87.4, "engine_generated": 256, "prompt_ms": 2156.0, "decode_ms": 2927.5, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.753, "tokens_per_round_rule": "derived: rounds = generated - drafts accepted (each verify round emits its accepted drafts plus one token, generate.cpp:4090-4119); the serve loop prints no per-request round count"}, "telemetry": {"window": {"t0": 1791538366.49, "t1": 1791538371.599, "seconds": 5.108, "host_samples": 0, "gpu_samples": 5}, "board": {"joules": 1044.76, "mean_w": 261.19, "max_w": 306.05, "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": 48.0, "ssd_max_c": null}, "vram": {"max_used_mib": 23820.0, "baseline_mib": 1.0, "growth_mib": 23819.0, "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"}}, "engine_env": {"STRATA_NO_LARGEPAGES": "1"}}
{"row_id": "R-S139-IQ3-24G/speed/a1/run-002", "arm": "R-S139-IQ3-24G", "engine": "strata", "engine_commit": "6f32ec070f23ced9f50e704d854d775da52591ab", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/usr/bin/python3", "/workshop/bench-flashnext/strata-6f32ec07/serve/server.py", "--engine", "strata", "--config", "/workshop/card24-strata/results/R-S139-IQ3-24G/R-S139-IQ3-24G.run/strata-config.json", "--port", "withheld-2026-10-13"], "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": 98.678, "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": 98.6, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "agree at run time (text recount at summary)", "completion_tokens": 256, "against": {"engine_generated (the 'N generated' log line)": 256, "metrics output_tokens": 256}, "gaps": {"engine_generated (the 'N generated' log line)": 0, "metrics output_tokens": 0}, "unread": [], "tolerance": "3% of the reported count", "rule": "Strata: usage completion_tokens against the engine's 'N generated' line and /metrics output_tokens (a gap above 3 % is COUNT-MISMATCH), then against the kept text recounted on llama.cpp's tokenizer at summary time (the same 3 %)"}, "token_frames": 249, "content_frames": 248, "frames": 250, "thinking_seen": false, "threads": {"workers": 31, "host_thread": true, "line": "strata generate: 31 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2163.9, "prompt_tok_s_client": 2124.0, "ttft_ms": 1901.146, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1866.1, "prompt_per_token_ms": 0.462, "prompt_per_second": 2163.9, "predicted_n": 256, "predicted_ms": 2597.2, "predicted_per_token_ms": 10.145, "predicted_per_second": 98.6, "draft_n": 216, "draft_n_accepted": 161}, "short_completion": false, "cache_flag": null, "completion_sha256": "e7067aa729f08b7d3d1f0b84f769e4d8ac60b3b44869be54cba96ce61e6eb0e7", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9209999999999999, "hits": 131689, "lookups": 142967, "hit_rate_line": "present", "offloaded": 7753, "pcie_share_log": 0.051, "routed": 150720, "drafts_accepted": 161, "drafts_offered": 216, "engine_generated": 256, "engine_decode_tok_s": 98.6, "engine_prompt_tok_s": 2163.8, "reused": 0, "metrics": {"hit_rate": 0.921, "pcie_share": 0.051, "decode_tok_s": 98.6, "engine_generated": 256, "prompt_ms": 1866.1, "decode_ms": 2597.2, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.695, "tokens_per_round_rule": "derived: rounds = generated - drafts accepted (each verify round emits its accepted drafts plus one token, generate.cpp:4090-4119); the serve loop prints no per-request round count"}, "telemetry": {"window": {"t0": 1791538371.604, "t1": 1791538376.089, "seconds": 4.485, "host_samples": 0, "gpu_samples": 5}, "board": {"joules": 1163.28, "mean_w": 290.75, "max_w": 299.69, "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": 50.0, "ssd_max_c": null}, "vram": {"max_used_mib": 23820.0, "baseline_mib": 1.0, "growth_mib": 23819.0, "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"}}, "engine_env": {"STRATA_NO_LARGEPAGES": "1"}}
{"row_id": "R-S139-IQ3-24G/speed/a1/run-003", "arm": "R-S139-IQ3-24G", "engine": "strata", "engine_commit": "6f32ec070f23ced9f50e704d854d775da52591ab", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/usr/bin/python3", "/workshop/bench-flashnext/strata-6f32ec07/serve/server.py", "--engine", "strata", "--config", "/workshop/card24-strata/results/R-S139-IQ3-24G/R-S139-IQ3-24G.run/strata-config.json", "--port", "withheld-2026-10-13"], "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": 98.291, "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": 98.1, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "agree at run time (text recount at summary)", "completion_tokens": 256, "against": {"engine_generated (the 'N generated' log line)": 256, "metrics output_tokens": 256}, "gaps": {"engine_generated (the 'N generated' log line)": 0, "metrics output_tokens": 0}, "unread": [], "tolerance": "3% of the reported count", "rule": "Strata: usage completion_tokens against the engine's 'N generated' line and /metrics output_tokens (a gap above 3 % is COUNT-MISMATCH), then against the kept text recounted on llama.cpp's tokenizer at summary time (the same 3 %)"}, "token_frames": 249, "content_frames": 248, "frames": 250, "thinking_seen": false, "threads": {"workers": 31, "host_thread": true, "line": "strata generate: 31 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2143.1, "prompt_tok_s_client": 2102.8, "ttft_ms": 1920.339, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1884.2, "prompt_per_token_ms": 0.467, "prompt_per_second": 2143.1, "predicted_n": 256, "predicted_ms": 2609.9, "predicted_per_token_ms": 10.195, "predicted_per_second": 98.1, "draft_n": 220, "draft_n_accepted": 157}, "short_completion": false, "cache_flag": null, "completion_sha256": "4d1555371f9e1bd011bf1cd6e29c3678d352c8958c015cbf44ac3adf30b83e01", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.927, "hits": 136258, "lookups": 146976, "hit_rate_line": "present", "offloaded": 7104, "pcie_share_log": 0.046, "routed": 154080, "drafts_accepted": 157, "drafts_offered": 220, "engine_generated": 256, "engine_decode_tok_s": 98.1, "engine_prompt_tok_s": 2143.1, "reused": 0, "metrics": {"hit_rate": 0.927, "pcie_share": 0.046, "decode_tok_s": 98.1, "engine_generated": 256, "prompt_ms": 1884.2, "decode_ms": 2609.9, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.586, "tokens_per_round_rule": "derived: rounds = generated - drafts accepted (each verify round emits its accepted drafts plus one token, generate.cpp:4090-4119); the serve loop prints no per-request round count"}, "telemetry": {"window": {"t0": 1791538376.092, "t1": 1791538380.607, "seconds": 4.515, "host_samples": 0, "gpu_samples": 4}, "board": {"joules": 843.38, "mean_w": 281.12, "max_w": 298.29, "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": 50.0, "ssd_max_c": null}, "vram": {"max_used_mib": 23820.0, "baseline_mib": 1.0, "growth_mib": 23819.0, "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"}}, "engine_env": {"STRATA_NO_LARGEPAGES": "1"}}
{"row_id": "R-S139-IQ3-24G/speed/a1/run-004", "arm": "R-S139-IQ3-24G", "engine": "strata", "engine_commit": "6f32ec070f23ced9f50e704d854d775da52591ab", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/usr/bin/python3", "/workshop/bench-flashnext/strata-6f32ec07/serve/server.py", "--engine", "strata", "--config", "/workshop/card24-strata/results/R-S139-IQ3-24G/R-S139-IQ3-24G.run/strata-config.json", "--port", "withheld-2026-10-13"], "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": 92.571, "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": 92.4, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "agree at run time (text recount at summary)", "completion_tokens": 256, "against": {"engine_generated (the 'N generated' log line)": 256, "metrics output_tokens": 256}, "gaps": {"engine_generated (the 'N generated' log line)": 0, "metrics output_tokens": 0}, "unread": [], "tolerance": "3% of the reported count", "rule": "Strata: usage completion_tokens against the engine's 'N generated' line and /metrics output_tokens (a gap above 3 % is COUNT-MISMATCH), then against the kept text recounted on llama.cpp's tokenizer at summary time (the same 3 %)"}, "token_frames": 249, "content_frames": 248, "frames": 250, "thinking_seen": false, "threads": {"workers": 31, "host_thread": true, "line": "strata generate: 31 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2671.6, "prompt_tok_s_client": 2638.2, "ttft_ms": 12143.82, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 11992.0, "prompt_per_token_ms": 0.374, "prompt_per_second": 2671.6, "predicted_n": 256, "predicted_ms": 2770.7, "predicted_per_token_ms": 10.823, "predicted_per_second": 92.4, "draft_n": 220, "draft_n_accepted": 151}, "short_completion": false, "cache_flag": null, "completion_sha256": "e3cdde9fd2287da6e2ead8c329b64adbfc0655de9cf614bf07aec53e25b9827e", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.922, "hits": 136712, "lookups": 148292, "hit_rate_line": "present", "offloaded": 7708, "pcie_share_log": 0.049, "routed": 156000, "drafts_accepted": 151, "drafts_offered": 220, "engine_generated": 256, "engine_decode_tok_s": 92.4, "engine_prompt_tok_s": 2671.6, "reused": 0, "metrics": {"hit_rate": 0.922, "pcie_share": 0.049, "decode_tok_s": 92.4, "engine_generated": 256, "prompt_ms": 11992.0, "decode_ms": 2770.7, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.438, "tokens_per_round_rule": "derived: rounds = generated - drafts accepted (each verify round emits its accepted drafts plus one token, generate.cpp:4090-4119); the serve loop prints no per-request round count"}, "telemetry": {"window": {"t0": 1791538380.611, "t1": 1791538395.509, "seconds": 14.899, "host_samples": 0, "gpu_samples": 15}, "board": {"joules": 4078.85, "mean_w": 291.31, "max_w": 300.32, "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": 23820.0, "baseline_mib": 1.0, "growth_mib": 23819.0, "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"}}, "engine_env": {"STRATA_NO_LARGEPAGES": "1"}}
{"row_id": "R-S139-IQ3-24G/speed/a1/run-005", "arm": "R-S139-IQ3-24G", "engine": "strata", "engine_commit": "6f32ec070f23ced9f50e704d854d775da52591ab", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/usr/bin/python3", "/workshop/bench-flashnext/strata-6f32ec07/serve/server.py", "--engine", "strata", "--config", "/workshop/card24-strata/results/R-S139-IQ3-24G/R-S139-IQ3-24G.run/strata-config.json", "--port", "withheld-2026-10-13"], "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": 94.246, "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": 94.1, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "agree at run time (text recount at summary)", "completion_tokens": 256, "against": {"engine_generated (the 'N generated' log line)": 256, "metrics output_tokens": 256}, "gaps": {"engine_generated (the 'N generated' log line)": 0, "metrics output_tokens": 0}, "unread": [], "tolerance": "3% of the reported count", "rule": "Strata: usage completion_tokens against the engine's 'N generated' line and /metrics output_tokens (a gap above 3 % is COUNT-MISMATCH), then against the kept text recounted on llama.cpp's tokenizer at summary time (the same 3 %)"}, "token_frames": 249, "content_frames": 248, "frames": 250, "thinking_seen": false, "threads": {"workers": 31, "host_thread": true, "line": "strata generate: 31 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2678.0, "prompt_tok_s_client": 2645.6, "ttft_ms": 12110.033, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 11963.5, "prompt_per_token_ms": 0.373, "prompt_per_second": 2678.0, "predicted_n": 256, "predicted_ms": 2719.8, "predicted_per_token_ms": 10.624, "predicted_per_second": 94.1, "draft_n": 217, "draft_n_accepted": 150}, "short_completion": false, "cache_flag": null, "completion_sha256": "0835927446847135d5a67a7a871cb2ab54437bcc2da02993b80bde7f56e8b304", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.925, "hits": 137004, "lookups": 148044, "hit_rate_line": "present", "offloaded": 6996, "pcie_share_log": 0.045, "routed": 155040, "drafts_accepted": 150, "drafts_offered": 217, "engine_generated": 256, "engine_decode_tok_s": 94.1, "engine_prompt_tok_s": 2678.0, "reused": 0, "metrics": {"hit_rate": 0.925, "pcie_share": 0.045, "decode_tok_s": 94.1, "engine_generated": 256, "prompt_ms": 11963.5, "decode_ms": 2719.8, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.415, "tokens_per_round_rule": "derived: rounds = generated - drafts accepted (each verify round emits its accepted drafts plus one token, generate.cpp:4090-4119); the serve loop prints no per-request round count"}, "telemetry": {"window": {"t0": 1791538395.513, "t1": 1791538410.329, "seconds": 14.816, "host_samples": 0, "gpu_samples": 15}, "board": {"joules": 4111.52, "mean_w": 293.66, "max_w": 298.92, "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": 58.0, "ssd_max_c": null}, "vram": {"max_used_mib": 23820.0, "baseline_mib": 1.0, "growth_mib": 23819.0, "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"}}, "engine_env": {"STRATA_NO_LARGEPAGES": "1"}}
{"row_id": "R-S139-IQ3-24G/speed/a1/run-006", "arm": "R-S139-IQ3-24G", "engine": "strata", "engine_commit": "6f32ec070f23ced9f50e704d854d775da52591ab", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/usr/bin/python3", "/workshop/bench-flashnext/strata-6f32ec07/serve/server.py", "--engine", "strata", "--config", "/workshop/card24-strata/results/R-S139-IQ3-24G/R-S139-IQ3-24G.run/strata-config.json", "--port", "withheld-2026-10-13"], "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": 90.121, "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": 90.0, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "agree at run time (text recount at summary)", "completion_tokens": 256, "against": {"engine_generated (the 'N generated' log line)": 256, "metrics output_tokens": 256}, "gaps": {"engine_generated (the 'N generated' log line)": 0, "metrics output_tokens": 0}, "unread": [], "tolerance": "3% of the reported count", "rule": "Strata: usage completion_tokens against the engine's 'N generated' line and /metrics output_tokens (a gap above 3 % is COUNT-MISMATCH), then against the kept text recounted on llama.cpp's tokenizer at summary time (the same 3 %)"}, "token_frames": 248, "content_frames": 247, "frames": 249, "thinking_seen": false, "threads": {"workers": 31, "host_thread": true, "line": "strata generate: 31 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2660.1, "prompt_tok_s_client": 2627.7, "ttft_ms": 12192.186, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 12044.1, "prompt_per_token_ms": 0.376, "prompt_per_second": 2660.1, "predicted_n": 256, "predicted_ms": 2844.0, "predicted_per_token_ms": 11.109, "predicted_per_second": 90.0, "draft_n": 225, "draft_n_accepted": 165}, "short_completion": false, "cache_flag": null, "completion_sha256": "2e3e9ab640e6af21f489d1e0058ff11abecd67ce4b8f4bd5874b4957c3ced3c5", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.8959999999999999, "hits": 127730, "lookups": 142531, "hit_rate_line": "present", "offloaded": 10589, "pcie_share_log": 0.069, "routed": 153120, "drafts_accepted": 165, "drafts_offered": 225, "engine_generated": 256, "engine_decode_tok_s": 90.0, "engine_prompt_tok_s": 2660.0, "reused": 0, "metrics": {"hit_rate": 0.896, "pcie_share": 0.069, "decode_tok_s": 90.0, "engine_generated": 256, "prompt_ms": 12044.1, "decode_ms": 2844.0, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.813, "tokens_per_round_rule": "derived: rounds = generated - drafts accepted (each verify round emits its accepted drafts plus one token, generate.cpp:4090-4119); the serve loop prints no per-request round count"}, "telemetry": {"window": {"t0": 1791538410.333, "t1": 1791538425.355, "seconds": 15.022, "host_samples": 0, "gpu_samples": 15}, "board": {"joules": 4124.56, "mean_w": 294.57, "max_w": 299.73, "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": 58.0, "ssd_max_c": null}, "vram": {"max_used_mib": 23820.0, "baseline_mib": 1.0, "growth_mib": 23819.0, "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"}}, "engine_env": {"STRATA_NO_LARGEPAGES": "1"}}
{"row_id": "R-S139-IQ3-24G/speed/a1/run-007", "arm": "R-S139-IQ3-24G", "engine": "strata", "engine_commit": "6f32ec070f23ced9f50e704d854d775da52591ab", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/usr/bin/python3", "/workshop/bench-flashnext/strata-6f32ec07/serve/server.py", "--engine", "strata", "--config", "/workshop/card24-strata/results/R-S139-IQ3-24G/R-S139-IQ3-24G.run/strata-config.json", "--port", "withheld-2026-10-13"], "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": 79.741, "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": 79.7, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "agree at run time (text recount at summary)", "completion_tokens": 256, "against": {"engine_generated (the 'N generated' log line)": 256, "metrics output_tokens": 256}, "gaps": {"engine_generated (the 'N generated' log line)": 0, "metrics output_tokens": 0}, "unread": [], "tolerance": "3% of the reported count", "rule": "Strata: usage completion_tokens against the engine's 'N generated' line and /metrics output_tokens (a gap above 3 % is COUNT-MISMATCH), then against the kept text recounted on llama.cpp's tokenizer at summary time (the same 3 %)"}, "token_frames": 250, "content_frames": 249, "frames": 251, "thinking_seen": false, "threads": {"workers": 31, "host_thread": true, "line": "strata generate: 31 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2550.7, "prompt_tok_s_client": 2523.2, "ttft_ms": 50743.571, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 50196.8, "prompt_per_token_ms": 0.392, "prompt_per_second": 2550.7, "predicted_n": 256, "predicted_ms": 3212.1, "predicted_per_token_ms": 12.547, "predicted_per_second": 79.7, "draft_n": 236, "draft_n_accepted": 138}, "short_completion": false, "cache_flag": null, "completion_sha256": "760e42c29b4e16caca16f55b00f351709b5a37579f6bce5b5df54e18741cc6e1", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.922, "hits": 148729, "lookups": 161294, "hit_rate_line": "present", "offloaded": 8626, "pcie_share_log": 0.051, "routed": 169920, "drafts_accepted": 138, "drafts_offered": 236, "engine_generated": 256, "engine_decode_tok_s": 79.7, "engine_prompt_tok_s": 2550.7, "reused": 0, "metrics": {"hit_rate": 0.922, "pcie_share": 0.051, "decode_tok_s": 79.7, "engine_generated": 256, "prompt_ms": 50196.8, "decode_ms": 3212.1, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.169, "tokens_per_round_rule": "derived: rounds = generated - drafts accepted (each verify round emits its accepted drafts plus one token, generate.cpp:4090-4119); the serve loop prints no per-request round count"}, "telemetry": {"window": {"t0": 1791538425.361, "t1": 1791538479.304, "seconds": 53.942, "host_samples": 0, "gpu_samples": 54}, "board": {"joules": 15640.45, "mean_w": 295.08, "max_w": 305.08, "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": 60.0, "ssd_max_c": null}, "vram": {"max_used_mib": 23820.0, "baseline_mib": 1.0, "growth_mib": 23819.0, "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"}}, "engine_env": {"STRATA_NO_LARGEPAGES": "1"}}
{"row_id": "R-S139-IQ3-24G/speed/a1/run-008", "arm": "R-S139-IQ3-24G", "engine": "strata", "engine_commit": "6f32ec070f23ced9f50e704d854d775da52591ab", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/usr/bin/python3", "/workshop/bench-flashnext/strata-6f32ec07/serve/server.py", "--engine", "strata", "--config", "/workshop/card24-strata/results/R-S139-IQ3-24G/R-S139-IQ3-24G.run/strata-config.json", "--port", "withheld-2026-10-13"], "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": 81.452, "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": 81.3, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "agree at run time (text recount at summary)", "completion_tokens": 256, "against": {"engine_generated (the 'N generated' log line)": 256, "metrics output_tokens": 256}, "gaps": {"engine_generated (the 'N generated' log line)": 0, "metrics output_tokens": 0}, "unread": [], "tolerance": "3% of the reported count", "rule": "Strata: usage completion_tokens against the engine's 'N generated' line and /metrics output_tokens (a gap above 3 % is COUNT-MISMATCH), then against the kept text recounted on llama.cpp's tokenizer at summary time (the same 3 %)"}, "token_frames": 250, "content_frames": 249, "frames": 251, "thinking_seen": false, "threads": {"workers": 31, "host_thread": true, "line": "strata generate: 31 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2532.4, "prompt_tok_s_client": 2505.4, "ttft_ms": 51104.739, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 50559.0, "prompt_per_token_ms": 0.395, "prompt_per_second": 2532.4, "predicted_n": 256, "predicted_ms": 3146.9, "predicted_per_token_ms": 12.293, "predicted_per_second": 81.3, "draft_n": 239, "draft_n_accepted": 144}, "short_completion": false, "cache_flag": null, "completion_sha256": "76755fedbdbc263902ca4bace08aa81f3e9b4d01cc58d31a052fcdbd0b52a6d2", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.925, "hits": 148721, "lookups": 160740, "hit_rate_line": "present", "offloaded": 8220, "pcie_share_log": 0.049, "routed": 168960, "drafts_accepted": 144, "drafts_offered": 239, "engine_generated": 256, "engine_decode_tok_s": 81.4, "engine_prompt_tok_s": 2532.4, "reused": 0, "metrics": {"hit_rate": 0.925, "pcie_share": 0.049, "decode_tok_s": 81.3, "engine_generated": 256, "prompt_ms": 50559.0, "decode_ms": 3146.9, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.286, "tokens_per_round_rule": "derived: rounds = generated - drafts accepted (each verify round emits its accepted drafts plus one token, generate.cpp:4090-4119); the serve loop prints no per-request round count"}, "telemetry": {"window": {"t0": 1791538479.311, "t1": 1791538533.547, "seconds": 54.236, "host_samples": 0, "gpu_samples": 54}, "board": {"joules": 15629.24, "mean_w": 294.86, "max_w": 303.06, "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": 61.0, "ssd_max_c": null}, "vram": {"max_used_mib": 23820.0, "baseline_mib": 1.0, "growth_mib": 23819.0, "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"}}, "engine_env": {"STRATA_NO_LARGEPAGES": "1"}}
{"row_id": "R-S139-IQ3-24G/speed/a1/run-009", "arm": "R-S139-IQ3-24G", "engine": "strata", "engine_commit": "6f32ec070f23ced9f50e704d854d775da52591ab", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/usr/bin/python3", "/workshop/bench-flashnext/strata-6f32ec07/serve/server.py", "--engine", "strata", "--config", "/workshop/card24-strata/results/R-S139-IQ3-24G/R-S139-IQ3-24G.run/strata-config.json", "--port", "withheld-2026-10-13"], "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": 89.219, "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": 89.0, "decode_tok_s_withheld": null, "decode_tok_s_server_withheld": null, "count_check": {"verdict": "agree at run time (text recount at summary)", "completion_tokens": 256, "against": {"engine_generated (the 'N generated' log line)": 256, "metrics output_tokens": 256}, "gaps": {"engine_generated (the 'N generated' log line)": 0, "metrics output_tokens": 0}, "unread": [], "tolerance": "3% of the reported count", "rule": "Strata: usage completion_tokens against the engine's 'N generated' line and /metrics output_tokens (a gap above 3 % is COUNT-MISMATCH), then against the kept text recounted on llama.cpp's tokenizer at summary time (the same 3 %)"}, "token_frames": 252, "content_frames": 251, "frames": 253, "thinking_seen": false, "threads": {"workers": 31, "host_thread": true, "line": "strata generate: 31 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2535.0, "prompt_tok_s_client": 2508.1, "ttft_ms": 51050.146, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 50507.5, "prompt_per_token_ms": 0.394, "prompt_per_second": 2535.0, "predicted_n": 256, "predicted_ms": 2874.8, "predicted_per_token_ms": 11.23, "predicted_per_second": 89.0, "draft_n": 228, "draft_n_accepted": 152}, "short_completion": false, "cache_flag": null, "completion_sha256": "370283f208d0a02a44e19c958bccce39f06819a64d61200fb9dc8f5e9c323191", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9279999999999999, "hits": 141286, "lookups": 152211, "hit_rate_line": "present", "offloaded": 7149, "pcie_share_log": 0.045, "routed": 159360, "drafts_accepted": 152, "drafts_offered": 228, "engine_generated": 256, "engine_decode_tok_s": 89.0, "engine_prompt_tok_s": 2535.0, "reused": 0, "metrics": {"hit_rate": 0.928, "pcie_share": 0.045, "decode_tok_s": 89.0, "engine_generated": 256, "prompt_ms": 50507.5, "decode_ms": 2874.8, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.462, "tokens_per_round_rule": "derived: rounds = generated - drafts accepted (each verify round emits its accepted drafts plus one token, generate.cpp:4090-4119); the serve loop prints no per-request round count"}, "telemetry": {"window": {"t0": 1791538533.552, "t1": 1791538587.461, "seconds": 53.909, "host_samples": 0, "gpu_samples": 54}, "board": {"joules": 15658.28, "mean_w": 295.41, "max_w": 301.25, "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": 61.0, "ssd_max_c": null}, "vram": {"max_used_mib": 23820.0, "baseline_mib": 1.0, "growth_mib": 23819.0, "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"}}, "engine_env": {"STRATA_NO_LARGEPAGES": "1"}}