{"row_id": "S-NEW/speed/a1/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/a1/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": 72.716, "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": 72.5, "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": 2022.9, "prompt_tok_s_client": 1985.1, "ttft_ms": 2034.178, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1996.1, "prompt_per_token_ms": 0.494, "prompt_per_second": 2022.9, "predicted_n": 256, "predicted_ms": 3532.2, "predicted_per_token_ms": 13.798, "predicted_per_second": 72.5, "draft_n": 237, "draft_n_accepted": 160}, "short_completion": false, "cache_flag": null, "completion_sha256": "b4d80f1fdd4bf968feee7d9fd1ce54f7f193f5509b5638ca5a3ddf13d8f7b090", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.873, "hits": 127882, "lookups": 146540, "hit_rate_line": "present", "offloaded": 13300, "pcie_share_log": 0.083, "routed": 159840, "drafts_accepted": 160, "drafts_offered": 237, "engine_generated": 256, "engine_decode_tok_s": 72.5, "engine_prompt_tok_s": 2023.0, "reused": 0, "metrics": {"hit_rate": 0.873, "pcie_share": 0.083, "decode_tok_s": 72.5, "engine_generated": 256, "prompt_ms": 1996.1, "decode_ms": 3532.2, "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": 1791214549.502, "t1": 1791214555.043, "seconds": 5.541, "host_samples": 6, "gpu_samples": 5}, "board": {"joules": 494.92, "mean_w": 123.67, "max_w": 165.57, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 493.38, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 5.003}, "data_reads_per_s": {"mean": 8542.9, "max": 47637.8}, "data_read_mb_per_s": {"mean": 36.22}, "temps": {"gpu_max_c": 69.0, "ssd_max_c": 57.85}, "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": 53337755648, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 7.11, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 2.656, "user_psi_full_avg10_max": 4.65, "psi_full_avg10_max": 5.23, "unit_psi_full_avg10_max": 1.45, "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": 148.18, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}, "engine_env": {"STRATA_NO_LARGEPAGES": "1"}}
{"row_id": "S-NEW/speed/a1/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/a1/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": 81.595, "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": 81.6, "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": 23, "host_thread": true, "line": "strata generate: 23 expert-pool workers + the host thread"}, "prompt_tok_s_server": 2290.2, "prompt_tok_s_client": 2249.4, "ttft_ms": 1795.108, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1763.2, "prompt_per_token_ms": 0.437, "prompt_per_second": 2290.2, "predicted_n": 256, "predicted_ms": 3138.4, "predicted_per_token_ms": 12.259, "predicted_per_second": 81.6, "draft_n": 210, "draft_n_accepted": 160}, "short_completion": false, "cache_flag": null, "completion_sha256": "bce216c1ccea865f706917d7383aa4ea6dac6032b3b7d8f40147a61db434547c", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9229999999999999, "hits": 128923, "lookups": 139655, "hit_rate_line": "present", "offloaded": 7225, "pcie_share_log": 0.049, "routed": 146880, "drafts_accepted": 160, "drafts_offered": 210, "engine_generated": 256, "engine_decode_tok_s": 81.6, "engine_prompt_tok_s": 2290.2, "reused": 0, "metrics": {"hit_rate": 0.923, "pcie_share": 0.049, "decode_tok_s": 81.6, "engine_generated": 256, "prompt_ms": 1763.2, "decode_ms": 3138.4, "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": 1791214555.046, "t1": 1791214559.968, "seconds": 4.922, "host_samples": 5, "gpu_samples": 5}, "board": {"joules": 516.54, "mean_w": 129.11, "max_w": 148.14, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 389.42, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 4.005}, "data_reads_per_s": {"mean": 427.0, "max": 762.3}, "data_read_mb_per_s": {"mean": 1.87}, "temps": {"gpu_max_c": 71.0, "ssd_max_c": 56.85}, "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": 53344821248, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 9.15, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 2.453, "user_psi_full_avg10_max": 3.12, "psi_full_avg10_max": 3.51, "unit_psi_full_avg10_max": 0.8, "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": 136.14, "rule": "enforced.power.limit READ (Dynamic Boost floats it), never one number"}}, "engine_env": {"STRATA_NO_LARGEPAGES": "1"}}
{"row_id": "S-NEW/speed/a1/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/a1/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": 81.088, "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": 80.9, "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": 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": 2250.8, "prompt_tok_s_client": 2211.7, "ttft_ms": 1825.746, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1794.0, "prompt_per_token_ms": 0.444, "prompt_per_second": 2250.8, "predicted_n": 256, "predicted_ms": 3164.3, "predicted_per_token_ms": 12.361, "predicted_per_second": 80.9, "draft_n": 226, "draft_n_accepted": 161}, "short_completion": false, "cache_flag": null, "completion_sha256": "dd8c25032b410d8e96d70f1e44e1c08610c0eb6ec83d0703e5aed3e61b8d289f", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": 0.9229999999999999, "hits": 135658, "lookups": 147005, "hit_rate_line": "present", "offloaded": 7555, "pcie_share_log": 0.049, "routed": 154560, "drafts_accepted": 161, "drafts_offered": 226, "engine_generated": 256, "engine_decode_tok_s": 80.9, "engine_prompt_tok_s": 2250.8, "reused": 0, "metrics": {"hit_rate": 0.923, "pcie_share": 0.049, "decode_tok_s": 80.9, "engine_generated": 256, "prompt_ms": 1794.0, "decode_ms": 3164.3, "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": 1791214559.971, "t1": 1791214564.942, "seconds": 4.971, "host_samples": 5, "gpu_samples": 5}, "board": {"joules": 497.69, "mean_w": 124.17, "max_w": 133.79, "rule": "trapezoid over nvidia-smi power.draw at 1 Hz"}, "rapl_package0": {"joules": 385.79, "label": "RAPL package-0, the CPU package: NOT WALL", "source": "/run/rapl/energy", "span_s": 4.006}, "data_reads_per_s": {"mean": 224.9, "max": 651.0}, "data_read_mb_per_s": {"mean": 0.98}, "temps": {"gpu_max_c": 73.0, "ssd_max_c": 55.85}, "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": 53344821248, "rule": "sums over the engine unit's processes' /proc/<pid>/status, every 5 s"}, "load": {"load1_mean": 11.15, "claude_procs_max": 1}, "memory": {"mem_available_gib_min": 2.717, "user_psi_full_avg10_max": 2.09, "psi_full_avg10_max": 3.17, "unit_psi_full_avg10_max": 0.53, "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"}}