{"row_id": "S-NEW/speed/a2/run-001", "arm": "S-NEW", "engine": "strata", "engine_commit": "6f32ec070f23ced9f50e704d854d775da52591ab", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-6f32ec07/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-NEW/speed/a2/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": 67.892, "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": 67.7, "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": 249, "content_frames": 248, "frames": 250, "thinking_seen": false, "threads": {"workers": 23, "host_thread": true, "line": "strata generate: 23 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2054.1, "prompt_tok_s_client": 2016.7, "ttft_ms": 2002.305, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1965.8, "prompt_per_token_ms": 0.487, "prompt_per_second": 2054.1, "predicted_n": 256, "predicted_ms": 3780.8, "predicted_per_token_ms": 14.769, "predicted_per_second": 67.7, "draft_n": 232, "draft_n_accepted": 160}, "short_completion": false, "cache_flag": null, "completion_sha256": "7bd5a83d3ff57476615ba117174b558f676f6588ad5ca34c7431bb1b1979df5f", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.875, "hits": 126451, "lookups": 144546, "hit_rate_line": "present", "offloaded": 12894, "pcie_share_log": 0.08199999999999999, "routed": 157440, "drafts_accepted": 160, "drafts_offered": 232, "engine_generated": 256, "engine_decode_tok_s": 67.7, "engine_prompt_tok_s": 2054.1, "reused": 0, "metrics": {"hit_rate": 0.875, "pcie_share": 0.082, "decode_tok_s": 67.7, "engine_generated": 256, "prompt_ms": 1965.8, "decode_ms": 3780.8, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.667, "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": 1791235189.906, "t1": 1791235195.664, "seconds": 5.758, "host_samples": 6, "gpu_samples": 6}, "board": {"joules": 497.87, "mean_w": 99.54, "max_w": 147.69, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 442.3, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 5.003}, "data_reads_per_s": {"mean": 8498.3, "max": 47657.8}, "data_read_mb_per_s": {"mean": 36.01}, "temps": {"gpu_max_c": 57.0, "ssd_max_c": 53.85}, "vram": {"max_used_mib": 23749.0, "baseline_mib": 15.0, "growth_mib": 23734.0, "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": 53391544320, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 3.03, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 2.203, "user_psi_full_avg10_max": 6.48, "psi_full_avg10_max": 7.98, "unit_psi_full_avg10_max": 1.72, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 0}, "power_limit": {"min_w": 135.0, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}, "engine_env": {"STRATA_NO_LARGEPAGES": "1"}}
{"row_id": "S-NEW/speed/a2/run-002", "arm": "S-NEW", "engine": "strata", "engine_commit": "6f32ec070f23ced9f50e704d854d775da52591ab", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-6f32ec07/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-NEW/speed/a2/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": 84.913, "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": 84.9, "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": 23, "host_thread": true, "line": "strata generate: 23 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2329.4, "prompt_tok_s_client": 2288.5, "ttft_ms": 1764.496, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1733.5, "prompt_per_token_ms": 0.429, "prompt_per_second": 2329.4, "predicted_n": 256, "predicted_ms": 3014.9, "predicted_per_token_ms": 11.777, "predicted_per_second": 84.9, "draft_n": 218, "draft_n_accepted": 154}, "short_completion": false, "cache_flag": null, "completion_sha256": "d600da3195980c4222e3448ab35f9a28371544692ae78ca372ef2623e9fd41c1", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.922, "hits": 134532, "lookups": 145941, "hit_rate_line": "present", "offloaded": 7659, "pcie_share_log": 0.05, "routed": 153600, "drafts_accepted": 154, "drafts_offered": 218, "engine_generated": 256, "engine_decode_tok_s": 84.9, "engine_prompt_tok_s": 2329.4, "reused": 0, "metrics": {"hit_rate": 0.922, "pcie_share": 0.05, "decode_tok_s": 84.9, "engine_generated": 256, "prompt_ms": 1733.5, "decode_ms": 3014.9, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.51, "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": 1791235195.669, "t1": 1791235200.436, "seconds": 4.768, "host_samples": 5, "gpu_samples": 5}, "board": {"joules": 504.99, "mean_w": 126.2, "max_w": 135.39, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 242.25, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 4.003}, "data_reads_per_s": {"mean": 543.8, "max": 1040.0}, "data_read_mb_per_s": {"mean": 2.27}, "temps": {"gpu_max_c": 57.0, "ssd_max_c": 52.85}, "vram": {"max_used_mib": 23749.0, "baseline_mib": 15.0, "growth_mib": 23734.0, "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": 53414141952, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 5.41, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 2.183, "user_psi_full_avg10_max": 4.28, "psi_full_avg10_max": 5.35, "unit_psi_full_avg10_max": 0.94, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 0}, "power_limit": {"min_w": 135.0, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}, "engine_env": {"STRATA_NO_LARGEPAGES": "1"}}
{"row_id": "S-NEW/speed/a2/run-003", "arm": "S-NEW", "engine": "strata", "engine_commit": "6f32ec070f23ced9f50e704d854d775da52591ab", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-6f32ec07/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-NEW/speed/a2/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": 92.444, "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": 92.3, "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": 249, "content_frames": 248, "frames": 250, "thinking_seen": false, "threads": {"workers": 23, "host_thread": true, "line": "strata generate: 23 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2213.6, "prompt_tok_s_client": 2149.4, "ttft_ms": 1878.638, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1824.2, "prompt_per_token_ms": 0.452, "prompt_per_second": 2213.6, "predicted_n": 256, "predicted_ms": 2774.9, "predicted_per_token_ms": 10.839, "predicted_per_second": 92.3, "draft_n": 224, "draft_n_accepted": 163}, "short_completion": false, "cache_flag": null, "completion_sha256": "6b1ce0bf869df6b6ca7747aa1447a1dbbf61da8ad7dd9e5974c6940caff6a160", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.92, "hits": 133521, "lookups": 145173, "hit_rate_line": "present", "offloaded": 7947, "pcie_share_log": 0.052000000000000005, "routed": 153120, "drafts_accepted": 163, "drafts_offered": 224, "engine_generated": 256, "engine_decode_tok_s": 92.3, "engine_prompt_tok_s": 2213.5, "reused": 0, "metrics": {"hit_rate": 0.92, "pcie_share": 0.052, "decode_tok_s": 92.3, "engine_generated": 256, "prompt_ms": 1824.2, "decode_ms": 2774.9, "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": 1791235200.439, "t1": 1791235205.077, "seconds": 4.637, "host_samples": 4, "gpu_samples": 5}, "board": {"joules": 494.5, "mean_w": 123.55, "max_w": 138.98, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 279.04, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 3.002}, "data_reads_per_s": {"mean": 141.9, "max": 274.0}, "data_read_mb_per_s": {"mean": 0.77}, "temps": {"gpu_max_c": 58.0, "ssd_max_c": 51.85}, "vram": {"max_used_mib": 23749.0, "baseline_mib": 15.0, "growth_mib": 23734.0, "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": 53414141952, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 6.9, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 2.454, "user_psi_full_avg10_max": 2.35, "psi_full_avg10_max": 4.45, "unit_psi_full_avg10_max": 0.51, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 0}, "power_limit": {"min_w": 135.0, "max_w": 135.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}, "engine_env": {"STRATA_NO_LARGEPAGES": "1"}}
{"row_id": "S-NEW/speed/a2/run-004", "arm": "S-NEW", "engine": "strata", "engine_commit": "6f32ec070f23ced9f50e704d854d775da52591ab", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-6f32ec07/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-NEW/speed/a2/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": 77.443, "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": 77.4, "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": 248, "content_frames": 247, "frames": 249, "thinking_seen": false, "threads": {"workers": 23, "host_thread": true, "line": "strata generate: 23 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2881.0, "prompt_tok_s_client": 2857.7, "ttft_ms": 11211.096, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 11120.4, "prompt_per_token_ms": 0.347, "prompt_per_second": 2881.0, "predicted_n": 256, "predicted_ms": 3308.9, "predicted_per_token_ms": 12.925, "predicted_per_second": 77.4, "draft_n": 241, "draft_n_accepted": 145}, "short_completion": false, "cache_flag": null, "completion_sha256": "4359e1daca574eb48af06e3540d597700a54d6cc231cdeff6b122426cb4584ce", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.929, "hits": 150355, "lookups": 161796, "hit_rate_line": "present", "offloaded": 7644, "pcie_share_log": 0.045, "routed": 169440, "drafts_accepted": 145, "drafts_offered": 241, "engine_generated": 256, "engine_decode_tok_s": 77.4, "engine_prompt_tok_s": 2881.0, "reused": 0, "metrics": {"hit_rate": 0.929, "pcie_share": 0.045, "decode_tok_s": 77.4, "engine_generated": 256, "prompt_ms": 11120.4, "decode_ms": 3308.9, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.306, "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": 1791235205.08, "t1": 1791235219.584, "seconds": 14.504, "host_samples": 15, "gpu_samples": 14}, "board": {"joules": 1815.05, "mean_w": 139.57, "max_w": 148.91, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 949.95, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 14.006}, "data_reads_per_s": {"mean": 11624.0, "max": 87334.0}, "data_read_mb_per_s": {"mean": 48.71}, "temps": {"gpu_max_c": 64.0, "ssd_max_c": 51.85}, "vram": {"max_used_mib": 23749.0, "baseline_mib": 15.0, "growth_mib": 23734.0, "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": 53481672704, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 7.95, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 2.425, "user_psi_full_avg10_max": 1.57, "psi_full_avg10_max": 3.64, "unit_psi_full_avg10_max": 0.34, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 0}, "power_limit": {"min_w": 135.0, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}, "engine_env": {"STRATA_NO_LARGEPAGES": "1"}}
{"row_id": "S-NEW/speed/a2/run-005", "arm": "S-NEW", "engine": "strata", "engine_commit": "6f32ec070f23ced9f50e704d854d775da52591ab", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-6f32ec07/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-NEW/speed/a2/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": 85.765, "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": 85.7, "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": 251, "content_frames": 250, "frames": 252, "thinking_seen": false, "threads": {"workers": 23, "host_thread": true, "line": "strata generate: 23 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2921.6, "prompt_tok_s_client": 2902.4, "ttft_ms": 11038.264, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 10965.8, "prompt_per_token_ms": 0.342, "prompt_per_second": 2921.6, "predicted_n": 256, "predicted_ms": 2988.9, "predicted_per_token_ms": 11.675, "predicted_per_second": 85.7, "draft_n": 228, "draft_n_accepted": 146}, "short_completion": false, "cache_flag": null, "completion_sha256": "691d31e0a4848c4d696b27537ce2671e08392d651c72454326f86119cad3c821", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9309999999999999, "hits": 144545, "lookups": 155218, "hit_rate_line": "present", "offloaded": 7022, "pcie_share_log": 0.043, "routed": 162240, "drafts_accepted": 146, "drafts_offered": 228, "engine_generated": 256, "engine_decode_tok_s": 85.6, "engine_prompt_tok_s": 2921.6, "reused": 0, "metrics": {"hit_rate": 0.931, "pcie_share": 0.043, "decode_tok_s": 85.7, "engine_generated": 256, "prompt_ms": 10965.8, "decode_ms": 2988.9, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.327, "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": 1791235219.586, "t1": 1791235233.597, "seconds": 14.012, "host_samples": 14, "gpu_samples": 14}, "board": {"joules": 1846.23, "mean_w": 141.94, "max_w": 150.3, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 607.74, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 13.003}, "data_reads_per_s": {"mean": 192.7, "max": 847.2}, "data_read_mb_per_s": {"mean": 0.81}, "temps": {"gpu_max_c": 66.0, "ssd_max_c": 50.85}, "vram": {"max_used_mib": 23749.0, "baseline_mib": 15.0, "growth_mib": 23734.0, "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": 53487226880, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 8.62, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 2.357, "user_psi_full_avg10_max": 2.97, "psi_full_avg10_max": 3.28, "unit_psi_full_avg10_max": 0.08, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 0}, "power_limit": {"min_w": 135.0, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}, "engine_env": {"STRATA_NO_LARGEPAGES": "1"}}
{"row_id": "S-NEW/speed/a2/run-006", "arm": "S-NEW", "engine": "strata", "engine_commit": "6f32ec070f23ced9f50e704d854d775da52591ab", "quant": "IQ3_S", "file_sha256": {"1": "4c1eb2ceb4915e1192f4f386021897bde56a97f40a0bb78bb86465e0f7d2aca3", "2": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113"}, "command_line": ["/data/strata-bench/venv/bin/python3", "/data/strata-bench/src/strata-6f32ec07/serve/server.py", "--engine", "strata", "--config", "/workshop/bench-archive/plans-2026-10-02/strata-bench/records/S-NEW/speed/a2/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": 88.005, "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": 87.9, "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": 251, "content_frames": 250, "frames": 252, "thinking_seen": false, "threads": {"workers": 23, "host_thread": true, "line": "strata generate: 23 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2923.9, "prompt_tok_s_client": 2904.6, "ttft_ms": 11030.126, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 10957.4, "prompt_per_token_ms": 0.342, "prompt_per_second": 2923.9, "predicted_n": 256, "predicted_ms": 2913.2, "predicted_per_token_ms": 11.38, "predicted_per_second": 87.9, "draft_n": 217, "draft_n_accepted": 156}, "short_completion": false, "cache_flag": null, "completion_sha256": "88b387c102c22ea69a41c8e66b905ed6121725e2fe5e4df08d0727b86357428c", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.93, "hits": 135882, "lookups": 146033, "hit_rate_line": "present", "offloaded": 6607, "pcie_share_log": 0.043, "routed": 152640, "drafts_accepted": 156, "drafts_offered": 217, "engine_generated": 256, "engine_decode_tok_s": 87.9, "engine_prompt_tok_s": 2923.9, "reused": 0, "metrics": {"hit_rate": 0.93, "pcie_share": 0.043, "decode_tok_s": 87.9, "engine_generated": 256, "prompt_ms": 10957.4, "decode_ms": 2913.2, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.56, "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": 1791235233.6, "t1": 1791235247.528, "seconds": 13.928, "host_samples": 14, "gpu_samples": 14}, "board": {"joules": 1851.32, "mean_w": 142.37, "max_w": 152.37, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 694.09, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 13.002}, "data_reads_per_s": {"mean": 110.6, "max": 616.9}, "data_read_mb_per_s": {"mean": 0.51}, "temps": {"gpu_max_c": 67.0, "ssd_max_c": 51.85}, "vram": {"max_used_mib": 23749.0, "baseline_mib": 15.0, "growth_mib": 23734.0, "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": 53487226880, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 7.61, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 2.272, "user_psi_full_avg10_max": 1.09, "psi_full_avg10_max": 1.2, "unit_psi_full_avg10_max": 0.01, "psi_rule": "psi = system-wide /proc/pressure/memory; user_psi = user@<uid>.service memory.pressure (systemd-oomd's); the stop reads the larger of the two; unit_psi = the engine unit's own memory.pressure (a figure, printed in every memory STOP)", "unit_max_events_delta": 0}, "power_limit": {"min_w": 135.0, "max_w": 150.0, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}, "engine_env": {"STRATA_NO_LARGEPAGES": "1"}}