{"row_id": "R-S139-IQ3/speed/a1/run-001", "arm": "R-S139-IQ3", "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/desktop-window-2/row/R-S139-IQ3.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": 202.136, "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": 199.8, "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": 2892.1, "prompt_tok_s_client": 2804.9, "ttft_ms": 1439.613, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1396.2, "prompt_per_token_ms": 0.346, "prompt_per_second": 2892.1, "predicted_n": 256, "predicted_ms": 1281.2, "predicted_per_token_ms": 5.005, "predicted_per_second": 199.8, "draft_n": 234, "draft_n_accepted": 163}, "short_completion": false, "cache_flag": null, "completion_sha256": "cca74af3b83cf5cbadf0536f9c11c2b8691457f74871c02cf3a63d9c3f2fb665", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": null, "hits": null, "lookups": null, "hit_rate_line": "absent", "offloaded": null, "pcie_share_log": null, "routed": null, "drafts_accepted": 163, "drafts_offered": 234, "engine_generated": 256, "engine_decode_tok_s": 199.8, "engine_prompt_tok_s": 2892.1, "reused": 0, "metrics": {"hit_rate": null, "pcie_share": null, "decode_tok_s": 199.8, "engine_generated": 256, "prompt_ms": 1396.2, "decode_ms": 1281.2, "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": 1791358742.505, "t1": 1791358745.206, "seconds": 2.701, "host_samples": 0, "gpu_samples": 3}, "board": {"joules": 401.32, "mean_w": 195.29, "max_w": 278.67, "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": 42.0, "ssd_max_c": null}, "vram": {"max_used_mib": 56388.0, "baseline_mib": 2.0, "growth_mib": 56386.0, "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/speed/a1/run-002", "arm": "R-S139-IQ3", "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/desktop-window-2/row/R-S139-IQ3.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": 205.263, "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": 204.9, "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": 31, "host_thread": true, "line": "strata generate: 31 expert-pool workers + the host thread"}, "prompt_tok_s_server": 3924.6, "prompt_tok_s_client": 3814.1, "ttft_ms": 1058.692, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1028.9, "prompt_per_token_ms": 0.255, "prompt_per_second": 3924.6, "predicted_n": 256, "predicted_ms": 1249.3, "predicted_per_token_ms": 4.88, "predicted_per_second": 204.9, "draft_n": 224, "draft_n_accepted": 158}, "short_completion": false, "cache_flag": null, "completion_sha256": "2604dcb319ad1631c0c44d393f861ca626306af98f24144af698b43f409b9e05", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": null, "hits": null, "lookups": null, "hit_rate_line": "absent", "offloaded": null, "pcie_share_log": null, "routed": null, "drafts_accepted": 158, "drafts_offered": 224, "engine_generated": 256, "engine_decode_tok_s": 204.9, "engine_prompt_tok_s": 3924.4, "reused": 0, "metrics": {"hit_rate": null, "pcie_share": null, "decode_tok_s": 204.9, "engine_generated": 256, "prompt_ms": 1028.9, "decode_ms": 1249.3, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.612, "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": 1791358745.213, "t1": 1791358747.514, "seconds": 2.301, "host_samples": 0, "gpu_samples": 2}, "board": {"joules": 298.71, "mean_w": 290.57, "max_w": 300.12, "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": 44.0, "ssd_max_c": null}, "vram": {"max_used_mib": 56388.0, "baseline_mib": 2.0, "growth_mib": 56386.0, "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/speed/a1/run-003", "arm": "R-S139-IQ3", "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/desktop-window-2/row/R-S139-IQ3.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": 212.961, "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": 212.6, "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": 3920.0, "prompt_tok_s_client": 3808.7, "ttft_ms": 1060.201, "prompt_tokens": 4038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 4038, "prompt_ms": 1030.1, "prompt_per_token_ms": 0.255, "prompt_per_second": 3920.0, "predicted_n": 256, "predicted_ms": 1204.3, "predicted_per_token_ms": 4.704, "predicted_per_second": 212.6, "draft_n": 221, "draft_n_accepted": 165}, "short_completion": false, "cache_flag": null, "completion_sha256": "f178e98544cb93b425882ac8b1438b5968e38a7b64699beb576bb32fe60ca258", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": null, "hits": null, "lookups": null, "hit_rate_line": "absent", "offloaded": null, "pcie_share_log": null, "routed": null, "drafts_accepted": 165, "drafts_offered": 221, "engine_generated": 256, "engine_decode_tok_s": 212.6, "engine_prompt_tok_s": 3920.0, "reused": 0, "metrics": {"hit_rate": null, "pcie_share": null, "decode_tok_s": 212.6, "engine_generated": 256, "prompt_ms": 1030.1, "decode_ms": 1204.3, "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": 1791358747.521, "t1": 1791358749.779, "seconds": 2.258, "host_samples": 0, "gpu_samples": 2}, "board": {"joules": 302.08, "mean_w": 293.86, "max_w": 299.88, "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": 44.0, "ssd_max_c": null}, "vram": {"max_used_mib": 56388.0, "baseline_mib": 2.0, "growth_mib": 56386.0, "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/speed/a1/run-004", "arm": "R-S139-IQ3", "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/desktop-window-2/row/R-S139-IQ3.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": 181.856, "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": 181.7, "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": 5033.9, "prompt_tok_s_client": 4922.9, "ttft_ms": 6507.943, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 6364.5, "prompt_per_token_ms": 0.199, "prompt_per_second": 5033.9, "predicted_n": 256, "predicted_ms": 1409.2, "predicted_per_token_ms": 5.505, "predicted_per_second": 181.7, "draft_n": 231, "draft_n_accepted": 147}, "short_completion": false, "cache_flag": null, "completion_sha256": "fff295d5486a2200dfa073dd7c53ed25c555b6994dd2823065aff49965b2b7f4", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": null, "hits": null, "lookups": null, "hit_rate_line": "absent", "offloaded": null, "pcie_share_log": null, "routed": null, "drafts_accepted": 147, "drafts_offered": 231, "engine_generated": 256, "engine_decode_tok_s": 181.7, "engine_prompt_tok_s": 5033.9, "reused": 0, "metrics": {"hit_rate": null, "pcie_share": null, "decode_tok_s": 181.7, "engine_generated": 256, "prompt_ms": 6364.5, "decode_ms": 1409.2, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.349, "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": 1791358749.786, "t1": 1791358757.697, "seconds": 7.911, "host_samples": 0, "gpu_samples": 8}, "board": {"joules": 2097.02, "mean_w": 291.54, "max_w": 300.41, "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": 46.0, "ssd_max_c": null}, "vram": {"max_used_mib": 56388.0, "baseline_mib": 2.0, "growth_mib": 56386.0, "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/speed/a1/run-005", "arm": "R-S139-IQ3", "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/desktop-window-2/row/R-S139-IQ3.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": 189.155, "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": 188.9, "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": 31, "host_thread": true, "line": "strata generate: 31 expert-pool workers + the host thread"}, "prompt_tok_s_server": 5154.7, "prompt_tok_s_client": 5040.5, "ttft_ms": 6356.087, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 6215.3, "prompt_per_token_ms": 0.194, "prompt_per_second": 5154.7, "predicted_n": 256, "predicted_ms": 1355.2, "predicted_per_token_ms": 5.294, "predicted_per_second": 188.9, "draft_n": 221, "draft_n_accepted": 151}, "short_completion": false, "cache_flag": null, "completion_sha256": "ed3e3e3c4f386b37802b50d28d74eef56d4ab5f152dc98a337e6eeba5b13f4c0", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": null, "hits": null, "lookups": null, "hit_rate_line": "absent", "offloaded": null, "pcie_share_log": null, "routed": null, "drafts_accepted": 151, "drafts_offered": 221, "engine_generated": 256, "engine_decode_tok_s": 188.9, "engine_prompt_tok_s": 5154.7, "reused": 0, "metrics": {"hit_rate": null, "pcie_share": null, "decode_tok_s": 188.9, "engine_generated": 256, "prompt_ms": 6215.3, "decode_ms": 1355.2, "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": 1791358757.704, "t1": 1791358765.408, "seconds": 7.704, "host_samples": 0, "gpu_samples": 7}, "board": {"joules": 1814.49, "mean_w": 294.27, "max_w": 304.01, "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": 47.0, "ssd_max_c": null}, "vram": {"max_used_mib": 56388.0, "baseline_mib": 2.0, "growth_mib": 56386.0, "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/speed/a1/run-006", "arm": "R-S139-IQ3", "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/desktop-window-2/row/R-S139-IQ3.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": 196.956, "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": 196.7, "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": 5143.9, "prompt_tok_s_client": 5029.6, "ttft_ms": 6369.871, "prompt_tokens": 32038, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 32038, "prompt_ms": 6228.3, "prompt_per_token_ms": 0.194, "prompt_per_second": 5143.9, "predicted_n": 256, "predicted_ms": 1301.7, "predicted_per_token_ms": 5.085, "predicted_per_second": 196.7, "draft_n": 225, "draft_n_accepted": 160}, "short_completion": false, "cache_flag": null, "completion_sha256": "8fa907db85bccf429f9ddd712bf77f7443f506200a305093c3601db2c7f1f06a", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": null, "hits": null, "lookups": null, "hit_rate_line": "absent", "offloaded": null, "pcie_share_log": null, "routed": null, "drafts_accepted": 160, "drafts_offered": 225, "engine_generated": 256, "engine_decode_tok_s": 196.7, "engine_prompt_tok_s": 5143.9, "reused": 0, "metrics": {"hit_rate": null, "pcie_share": null, "decode_tok_s": 196.7, "engine_generated": 256, "prompt_ms": 6228.3, "decode_ms": 1301.7, "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": 1791358765.416, "t1": 1791358773.081, "seconds": 7.665, "host_samples": 0, "gpu_samples": 8}, "board": {"joules": 2130.33, "mean_w": 296.21, "max_w": 300.78, "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": 56388.0, "baseline_mib": 2.0, "growth_mib": 56386.0, "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/speed/a1/run-007", "arm": "R-S139-IQ3", "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/desktop-window-2/row/R-S139-IQ3.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": 163.457, "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": 163.3, "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": 4980.4, "prompt_tok_s_client": 4877.6, "ttft_ms": 26250.25, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 25708.0, "prompt_per_token_ms": 0.201, "prompt_per_second": 4980.4, "predicted_n": 256, "predicted_ms": 1567.4, "predicted_per_token_ms": 6.123, "predicted_per_second": 163.3, "draft_n": 235, "draft_n_accepted": 138}, "short_completion": false, "cache_flag": null, "completion_sha256": "0b45a8cbd8743635083409f62d21dcd9563b30cbd0ad4feda31bf42df4ca58d3", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": null, "hits": null, "lookups": null, "hit_rate_line": "absent", "offloaded": null, "pcie_share_log": null, "routed": null, "drafts_accepted": 138, "drafts_offered": 235, "engine_generated": 256, "engine_decode_tok_s": 163.3, "engine_prompt_tok_s": 4980.4, "reused": 0, "metrics": {"hit_rate": null, "pcie_share": null, "decode_tok_s": 163.3, "engine_generated": 256, "prompt_ms": 25708.0, "decode_ms": 1567.4, "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": 1791358773.094, "t1": 1791358800.906, "seconds": 27.812, "host_samples": 0, "gpu_samples": 27}, "board": {"joules": 7861.63, "mean_w": 296.44, "max_w": 315.28, "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": 56388.0, "baseline_mib": 2.0, "growth_mib": 56386.0, "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/speed/a1/run-008", "arm": "R-S139-IQ3", "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/desktop-window-2/row/R-S139-IQ3.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": 172.325, "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": 172.1, "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": 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": 4908.1, "prompt_tok_s_client": 4808.2, "ttft_ms": 26628.855, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 26087.0, "prompt_per_token_ms": 0.204, "prompt_per_second": 4908.1, "predicted_n": 256, "predicted_ms": 1487.1, "predicted_per_token_ms": 5.809, "predicted_per_second": 172.1, "draft_n": 241, "draft_n_accepted": 149}, "short_completion": false, "cache_flag": null, "completion_sha256": "c3f48fb370b2aba5a4f14f83eb366604a543c9668d65f760eee327bc2078d527", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": null, "hits": null, "lookups": null, "hit_rate_line": "absent", "offloaded": null, "pcie_share_log": null, "routed": null, "drafts_accepted": 149, "drafts_offered": 241, "engine_generated": 256, "engine_decode_tok_s": 172.1, "engine_prompt_tok_s": 4908.1, "reused": 0, "metrics": {"hit_rate": null, "pcie_share": null, "decode_tok_s": 172.1, "engine_generated": 256, "prompt_ms": 26087.0, "decode_ms": 1487.1, "reused": 0, "output_tokens": 256, "finish": "length", "projection": null}, "tokens_per_round_derived": 2.393, "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": 1791358800.92, "t1": 1791358829.03, "seconds": 28.109, "host_samples": 0, "gpu_samples": 28}, "board": {"joules": 8140.33, "mean_w": 294.63, "max_w": 316.49, "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": 56388.0, "baseline_mib": 2.0, "growth_mib": 56386.0, "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/speed/a1/run-009", "arm": "R-S139-IQ3", "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/desktop-window-2/row/R-S139-IQ3.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": 178.13, "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": 177.9, "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": 31, "host_thread": true, "line": "strata generate: 31 expert-pool workers + the host thread"}, "prompt_tok_s_server": 4873.2, "prompt_tok_s_client": 4775.6, "ttft_ms": 26810.397, "prompt_tokens": 128037, "completion_tokens": 256, "cached_tokens": 0, "finish_reason": "length", "server_timings": {"cache_n": 0, "prompt_n": 128037, "prompt_ms": 26273.9, "prompt_per_token_ms": 0.205, "prompt_per_second": 4873.2, "predicted_n": 256, "predicted_ms": 1438.9, "predicted_per_token_ms": 5.621, "predicted_per_second": 177.9, "draft_n": 219, "draft_n_accepted": 150}, "short_completion": false, "cache_flag": null, "completion_sha256": "552b4f3902f46b3fa442b995a7a70b90ce9ac1a49dfdd109883f67a6ca1be6b3", "completion_head": "withheld-2026-10-13", "strata": {"hit_rate_log": null, "hits": null, "lookups": null, "hit_rate_line": "absent", "offloaded": null, "pcie_share_log": null, "routed": null, "drafts_accepted": 150, "drafts_offered": 219, "engine_generated": 256, "engine_decode_tok_s": 177.9, "engine_prompt_tok_s": 4873.2, "reused": 0, "metrics": {"hit_rate": null, "pcie_share": null, "decode_tok_s": 177.9, "engine_generated": 256, "prompt_ms": 26273.9, "decode_ms": 1438.9, "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": 1791358829.039, "t1": 1791358857.282, "seconds": 28.243, "host_samples": 0, "gpu_samples": 27}, "board": {"joules": 7908.46, "mean_w": 296.96, "max_w": 310.16, "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": 62.0, "ssd_max_c": null}, "vram": {"max_used_mib": 56388.0, "baseline_mib": 2.0, "growth_mib": 56386.0, "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"}}