spb/wp33_uqo

1{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:02:22", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 517.0783339999616, "prompt_tps": 35.20778092655101, "gen_tps": 33.4326627330368, "decode_tps_wall": 32.911857674670465, "itl_ms_p50": 30.212124926038086, "itl_ms_p95": 39.193958044052124, "itl_ms_p99": 40.8845420461148, "itl_ms_mean": 30.384185842223584, "total_s": 2.4312820420600474, "peak_mem_gb": 1.829203036, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.4996987346856799, "cache_misses": 4982, "cache_hits": 4976, "stall_s": 1.3353543087141588, "hits_device": 0, "layers_complete": 0, "prefetch": null, "host_time_s": {"sync": 0.9596480529289693, "prefetch": 0.0, "ensure": 1.3742589412722737, "build": 0.01710272755008191}, "ssd_bytes": 17631019008, "ssd_read_s": 3.426499897032045}2{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:02:25", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 348.18725008517504, "prompt_tps": 40.253284123634536, "gen_tps": 32.55307470525428, "decode_tps_wall": 32.04460133325126, "itl_ms_p50": 30.876416945829988, "itl_ms_p95": 39.19358307030052, "itl_ms_p99": 43.04183297790587, "itl_ms_mean": 31.206504633975406, "total_s": 2.3141970420256257, "peak_mem_gb": 1.747773596, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.4795243019648397, "cache_misses": 5033, "cache_hits": 4637, "stall_s": 1.2445547764655203, "hits_device": 0, "layers_complete": 0, "prefetch": null, "host_time_s": {"sync": 0.9447881018277258, "prefetch": 0.0, "ensure": 1.2828235330525786, "build": 0.016037469962611794}, "ssd_bytes": 17811505152, "ssd_read_s": 3.1572994944872335}3{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:02:27", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 378.3321250230074, "prompt_tps": 47.62164073300397, "gen_tps": 33.273626374779, "decode_tps_wall": 32.75394520951318, "itl_ms_p50": 29.696666984818876, "itl_ms_p95": 38.782707997597754, "itl_ms_p99": 40.73108301963657, "itl_ms_mean": 30.530673285414064, "total_s": 2.3017645420040935, "peak_mem_gb": 1.812821016, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.5051215103434424, "cache_misses": 4928, "cache_hits": 5030, "stall_s": 1.2373295988654718, "hits_device": 0, "layers_complete": 0, "prefetch": null, "host_time_s": {"sync": 0.9343643408501521, "prefetch": 0.0, "ensure": 1.2764358408749104, "build": 0.01720138662494719}, "ssd_bytes": 17439916032, "ssd_read_s": 3.0876483746105805}4{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:02:29", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 343.4790410101414, "prompt_tps": 40.801873390380386, "gen_tps": 32.914169802976595, "decode_tps_wall": 32.400085155163254, "itl_ms_p50": 30.86812503170222, "itl_ms_p95": 39.35583296697587, "itl_ms_p99": 43.849790934473276, "itl_ms_mean": 30.864116412380504, "total_s": 2.287918374990113, "peak_mem_gb": 1.747773596, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.4795243019648397, "cache_misses": 5033, "cache_hits": 4637, "stall_s": 1.2332852088147774, "hits_device": 0, "layers_complete": 0, "prefetch": null, "host_time_s": {"sync": 0.9296224870486185, "prefetch": 0.0, "ensure": 1.271775528904982, "build": 0.016064595663920045}, "ssd_bytes": 17811505152, "ssd_read_s": 3.1249096140963957}5{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:02:31", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 342.86870807409286, "prompt_tps": 40.86731643441979, "gen_tps": 32.636451601945915, "decode_tps_wall": 32.12670637617493, "itl_ms_p50": 31.254375004209578, "itl_ms_p95": 39.06466602347791, "itl_ms_p99": 41.918667033314705, "itl_ms_mean": 31.12675131682957, "total_s": 2.3038540410343558, "peak_mem_gb": 1.741482148, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.4792140641158221, "cache_misses": 5036, "cache_hits": 4634, "stall_s": 1.2375774497631937, "hits_device": 0, "layers_complete": 0, "prefetch": null, "host_time_s": {"sync": 0.9419836424058303, "prefetch": 0.0, "ensure": 1.2754085791530088, "build": 0.016049782163463533}, "ssd_bytes": 17822121984, "ssd_read_s": 3.1362346468959004}6{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:02:34", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 381.7877920810133, "prompt_tps": 47.18251065755433, "gen_tps": 33.0810152124823, "decode_tps_wall": 32.56430248431398, "itl_ms_p50": 29.244499979540706, "itl_ms_p95": 39.76695891469717, "itl_ms_p99": 42.16379206627607, "itl_ms_mean": 30.708472889345433, "total_s": 2.3164215841097757, "peak_mem_gb": 1.832006824, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.5051215103434424, "cache_misses": 4928, "cache_hits": 5030, "stall_s": 1.2385562813142315, "hits_device": 0, "layers_complete": 0, "prefetch": null, "host_time_s": {"sync": 0.9488227724796161, "prefetch": 0.0, "ensure": 1.2772458243416622, "build": 0.016906953300349414}, "ssd_bytes": 17439916032, "ssd_read_s": 3.0955757402116433}7{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:02:41", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 543.5582080390304, "prompt_tps": 33.29991009188411, "gen_tps": 28.225582238301115, "decode_tps_wall": 27.78593305891577, "itl_ms_p50": 36.18829196784645, "itl_ms_p95": 44.08766608685255, "itl_ms_p99": 49.420999945141375, "itl_ms_mean": 35.98943385776014, "total_s": 2.8108925410779193, "peak_mem_gb": 1.849931992, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.6504318136171922, "cache_misses": 3481, "cache_hits": 6477, "stall_s": 1.2660493151051924, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3315, "installed": 3313, "late": 110, "dup": 5333, "predictor": "transition2", "m": 8}, "host_time_s": {"sync": 1.3203314587008208, "prefetch": 0.1151198975276202, "ensure": 1.2682464926037937, "build": 0.019508084165863693}, "ssd_bytes": 24050663424, "ssd_read_s": 4.82381077145692}8{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:02:44", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 363.3501670556143, "prompt_tps": 38.570176025229074, "gen_tps": 26.121693316338536, "decode_tps_wall": 25.713670423271733, "itl_ms_p50": 38.63416600506753, "itl_ms_p95": 45.95149995293468, "itl_ms_p99": 52.62233305256814, "itl_ms_mean": 38.889819443861526, "total_s": 2.8134087920188904, "peak_mem_gb": 1.766271896, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.6188210961737332, "cache_misses": 3686, "cache_hits": 5984, "stall_s": 1.2340706296963617, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 4201, "installed": 4195, "late": 107, "dup": 4351, "predictor": "transition2", "m": 8}, "host_time_s": {"sync": 1.3461572570959106, "prefetch": 0.12368838163092732, "ensure": 1.2391801627818495, "build": 0.01933034381363541}, "ssd_bytes": 27911651328, "ssd_read_s": 5.115074231871404}9{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:02:46", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 395.5610409611836, "prompt_tps": 45.556328182596026, "gen_tps": 28.245109293687104, "decode_tps_wall": 27.803938979960005, "itl_ms_p50": 36.07900009956211, "itl_ms_p95": 42.57133405189961, "itl_ms_p99": 49.7471250128001, "itl_ms_mean": 35.96612698368965, "total_s": 2.6614270409336314, "peak_mem_gb": 1.864631956, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.6567583852179153, "cache_misses": 3418, "cache_hits": 6540, "stall_s": 1.1613253159448504, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3257, "installed": 3263, "late": 128, "dup": 5455, "predictor": "transition2", "m": 8}, "host_time_s": {"sync": 1.2785086863441393, "prefetch": 0.12367042782716453, "ensure": 1.1563438426237553, "build": 0.0191463737282902}, "ssd_bytes": 23622451200, "ssd_read_s": 4.365179446875118}10{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:02:49", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 360.393957933411, "prompt_tps": 38.880581717523405, "gen_tps": 26.03441502874237, "decode_tps_wall": 25.627780626094925, "itl_ms_p50": 39.00966595392674, "itl_ms_p95": 46.859833993948996, "itl_ms_p99": 49.87420805264264, "itl_ms_mean": 39.02015607944497, "total_s": 2.8186637909384444, "peak_mem_gb": 1.762732952, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.6191313340227508, "cache_misses": 3683, "cache_hits": 5987, "stall_s": 1.246894485782832, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 4204, "installed": 4200, "late": 116, "dup": 4348, "predictor": "transition2", "m": 8}, "host_time_s": {"sync": 1.3361439587315544, "prefetch": 0.1258306608069688, "ensure": 1.2519114535534754, "build": 0.019310060422867537}, "ssd_bytes": 27911651328, "ssd_read_s": 5.152700936654583}11{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:02:52", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 359.2003750381991, "prompt_tps": 39.013927651294985, "gen_tps": 26.024585035633084, "decode_tps_wall": 25.618071984511023, "itl_ms_p50": 39.175417041406035, "itl_ms_p95": 47.21666604746133, "itl_ms_p99": 49.61366706993431, "itl_ms_mean": 39.034943792983775, "total_s": 2.8184018339961767, "peak_mem_gb": 1.766140932, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.6191313340227508, "cache_misses": 3683, "cache_hits": 5987, "stall_s": 1.243524986319244, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 4199, "installed": 4199, "late": 110, "dup": 4353, "predictor": "transition2", "m": 8}, "host_time_s": {"sync": 1.3395317263202742, "prefetch": 0.12579591968096793, "ensure": 1.247619368485175, "build": 0.019359704107046127}, "ssd_bytes": 27893956608, "ssd_read_s": 5.161090494017117}12{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:02:55", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 392.5358339911327, "prompt_tps": 45.89750167336509, "gen_tps": 28.221729007624088, "decode_tps_wall": 27.780911509162685, "itl_ms_p50": 36.173832952044904, "itl_ms_p95": 44.07958302181214, "itl_ms_p99": 46.510166954249144, "itl_ms_mean": 35.995939142248105, "total_s": 2.6602799999527633, "peak_mem_gb": 1.842349792, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.6565575416750351, "cache_misses": 3420, "cache_hits": 6538, "stall_s": 1.16297057911288, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3258, "installed": 3260, "late": 118, "dup": 5454, "predictor": "transition2", "m": 8}, "host_time_s": {"sync": 1.2777759277960286, "prefetch": 0.11877476540394127, "ensure": 1.1620669818948954, "build": 0.01921978371683508}, "ssd_bytes": 23633068032, "ssd_read_s": 4.389560426585376}13{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2t50", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2t50 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:03:52", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 529.1794170625508, "prompt_tps": 34.192393477150624, "gen_tps": 29.86195784549739, "decode_tps_wall": 29.39630203157457, "itl_ms_p50": 34.54083402175456, "itl_ms_p95": 41.63816594518721, "itl_ms_p99": 46.15004197694361, "itl_ms_mean": 34.01788425380512, "total_s": 2.6723061250522733, "peak_mem_gb": 1.842854104, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.641594697730468, "cache_misses": 3569, "cache_hits": 6389, "stall_s": 1.1812305508647114, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2982, "installed": 2982, "late": 68, "dup": 4886, "predictor": "transition2t50", "m": 8}, "host_time_s": {"sync": 1.2894892855547369, "prefetch": 0.093443124438636, "ensure": 1.1950585045851767, "build": 0.01706448826007545}, "ssd_bytes": 23183622144, "ssd_read_s": 4.405051150824875}14{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2t50", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2t50 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:03:54", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 347.15437493287027, "prompt_tps": 40.366358260015424, "gen_tps": 27.84939114644391, "decode_tps_wall": 27.41434630785206, "itl_ms_p50": 36.282457993365824, "itl_ms_p95": 43.52620802819729, "itl_ms_p99": 46.807541977614164, "itl_ms_mean": 36.4772513183573, "total_s": 2.6452212079893798, "peak_mem_gb": 1.75919402, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.612306101344364, "cache_misses": 3749, "cache_hits": 5921, "stall_s": 1.1427990465890616, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3922, "installed": 3917, "late": 64, "dup": 4031, "predictor": "transition2t50", "m": 8}, "host_time_s": {"sync": 1.2900060514220968, "prefetch": 0.10061502200551331, "ensure": 1.1604060878744349, "build": 0.01761058170814067}, "ssd_bytes": 27147239424, "ssd_read_s": 4.670730734593235}15{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2t50", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2t50 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:03:57", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 371.8323749490082, "prompt_tps": 48.44667837686298, "gen_tps": 29.41238423146309, "decode_tps_wall": 28.952942132932755, "itl_ms_p50": 34.52545800246298, "itl_ms_p95": 41.58550000283867, "itl_ms_p99": 44.41204201430082, "itl_ms_mean": 34.53880422268181, "total_s": 2.5477770409779623, "peak_mem_gb": 1.85047618, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.6476200040168708, "cache_misses": 3509, "cache_hits": 6449, "stall_s": 1.0857952203368768, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2935, "installed": 2937, "late": 75, "dup": 4997, "predictor": "transition2t50", "m": 8}, "host_time_s": {"sync": 1.26257005892694, "prefetch": 0.09576272440608591, "ensure": 1.098084571887739, "build": 0.017263506189920008}, "ssd_bytes": 22804955136, "ssd_read_s": 4.011304949061014}16{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2t50", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2t50 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:03:59", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 338.45895901322365, "prompt_tps": 41.400091973073614, "gen_tps": 27.448277821131576, "decode_tps_wall": 27.01951678370005, "itl_ms_p50": 36.77250002510846, "itl_ms_p95": 43.173833983018994, "itl_ms_p99": 48.3146250480786, "itl_ms_mean": 37.010284380928155, "total_s": 2.6701068750116974, "peak_mem_gb": 1.75919402, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.6126163391933815, "cache_misses": 3746, "cache_hits": 5924, "stall_s": 1.1472254098625854, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3926, "installed": 3924, "late": 57, "dup": 4027, "predictor": "transition2t50", "m": 8}, "host_time_s": {"sync": 1.3082013352541253, "prefetch": 0.1002028794027865, "ensure": 1.166193835902959, "build": 0.01799292885698378}, "ssd_bytes": 27150778368, "ssd_read_s": 4.7154672087635845}17{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2t50", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2t50 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:04:02", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 344.41329108085483, "prompt_tps": 40.68289139770562, "gen_tps": 26.832091171562322, "decode_tps_wall": 26.412956008529626, "itl_ms_p50": 37.27545903529972, "itl_ms_p95": 45.126124983653426, "itl_ms_p99": 51.0092080803588, "itl_ms_mean": 37.86020768281546, "total_s": 2.7296063750982285, "peak_mem_gb": 1.776888728, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.6120992761116856, "cache_misses": 3751, "cache_hits": 5919, "stall_s": 1.167463383404538, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3923, "installed": 3922, "late": 66, "dup": 4030, "predictor": "transition2t50", "m": 8}, "host_time_s": {"sync": 1.3267823717324063, "prefetch": 0.11268862476572394, "ensure": 1.1873078583739698, "build": 0.01936343580018729}, "ssd_bytes": 27157856256, "ssd_read_s": 4.747054176754318}18{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2t50", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2t50 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:04:05", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 385.8761249575764, "prompt_tps": 46.6842459752629, "gen_tps": 28.40039964078331, "decode_tps_wall": 27.95676331223442, "itl_ms_p50": 36.158874980174005, "itl_ms_p95": 42.71891701500863, "itl_ms_p99": 46.233917004428804, "itl_ms_mean": 35.769519841461076, "total_s": 2.639355874969624, "peak_mem_gb": 1.83548528, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.6477204257883109, "cache_misses": 3508, "cache_hits": 6450, "stall_s": 1.1170780754182488, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2934, "installed": 2937, "late": 85, "dup": 4998, "predictor": "transition2t50", "m": 8}, "host_time_s": {"sync": 1.2881446726387367, "prefetch": 0.11748608236666769, "ensure": 1.1284848203649744, "build": 0.020568106789141893}, "ssd_bytes": 22797877248, "ssd_read_s": 4.073544128099456}19{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2t70", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2t70 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:04:12", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 520.4024999402463, "prompt_tps": 34.79124767995126, "gen_tps": 29.01467632571752, "decode_tps_wall": 28.562278602486078, "itl_ms_p50": 34.93270801845938, "itl_ms_p95": 43.403832940384746, "itl_ms_p99": 43.99354197084904, "itl_ms_mean": 35.0112123026823, "total_s": 2.7261088750092313, "peak_mem_gb": 1.83970846, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.5900783289817232, "cache_misses": 4082, "cache_hits": 5876, "stall_s": 1.2609232818940654, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1737, "installed": 1737, "late": 8, "dup": 3275, "predictor": "transition2t70", "m": 8}, "host_time_s": {"sync": 1.2682854104787111, "prefetch": 0.0691060513490811, "ensure": 1.2916151109384373, "build": 0.01749010500498116}, "ssd_bytes": 20593115136, "ssd_read_s": 4.132487543975003}20{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2t70", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2t70 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:04:14", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 363.68562490679324, "prompt_tps": 38.5290569285975, "gen_tps": 27.939468551384746, "decode_tps_wall": 27.50304593514092, "itl_ms_p50": 36.37795790564269, "itl_ms_p95": 43.44491695519537, "itl_ms_p99": 47.161499969661236, "itl_ms_mean": 36.359609126867284, "total_s": 2.654340999899432, "peak_mem_gb": 1.751334372, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.5699069286452947, "cache_misses": 4159, "cache_hits": 5511, "stall_s": 1.186734238639474, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2312, "installed": 2312, "late": 10, "dup": 2668, "predictor": "transition2t70", "m": 8}, "host_time_s": {"sync": 1.2761520012281835, "prefetch": 0.07300603704061359, "ensure": 1.2156566720223054, "build": 0.016580969910137355}, "ssd_bytes": 22900506624, "ssd_read_s": 4.225527738803066}21{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2t70", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2t70 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:04:17", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 375.0706249848008, "prompt_tps": 48.03028032344766, "gen_tps": 29.388223861291042, "decode_tps_wall": 28.92916348541071, "itl_ms_p50": 34.04141601640731, "itl_ms_p95": 42.17079107183963, "itl_ms_p99": 44.80775003321469, "itl_ms_mean": 34.567193776768235, "total_s": 2.5528038329211995, "peak_mem_gb": 1.823409816, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.5956015264109259, "cache_misses": 4027, "cache_hits": 5931, "stall_s": 1.1404099247884005, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1707, "installed": 1707, "late": 12, "dup": 3369, "predictor": "transition2t70", "m": 8}, "host_time_s": {"sync": 1.224560545757413, "prefetch": 0.06828455091454089, "ensure": 1.170226154383272, "build": 0.016853181179612875}, "ssd_bytes": 20292304896, "ssd_read_s": 3.6672123356256634}22{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2t70", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2t70 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:04:20", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 344.3355839699507, "prompt_tps": 40.69267655882825, "gen_tps": 28.418940554095958, "decode_tps_wall": 27.975041089827275, "itl_ms_p50": 35.822125035338104, "itl_ms_p95": 42.88266703952104, "itl_ms_p99": 46.16183403413743, "itl_ms_mean": 35.746149461908594, "total_s": 2.596343000070192, "peak_mem_gb": 1.734421476, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.569493278179938, "cache_misses": 4163, "cache_hits": 5507, "stall_s": 1.1820058053126559, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2316, "installed": 2316, "late": 12, "dup": 2664, "predictor": "transition2t70", "m": 8}, "host_time_s": {"sync": 1.226772032561712, "prefetch": 0.07226992608048022, "ensure": 1.210089394939132, "build": 0.015987955266609788}, "ssd_bytes": 22928818176, "ssd_read_s": 4.200161305372603}23{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2t70", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2t70 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:04:22", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 355.18995800521225, "prompt_tps": 39.448185592933086, "gen_tps": 27.225612276316657, "decode_tps_wall": 26.800377864463773, "itl_ms_p50": 37.167792092077434, "itl_ms_p95": 43.930709012784064, "itl_ms_p99": 47.71245794836432, "itl_ms_mean": 37.31290674546645, "total_s": 2.7059030829695985, "peak_mem_gb": 1.761951132, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.5695966907962772, "cache_misses": 4162, "cache_hits": 5508, "stall_s": 1.1938807795522735, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2315, "installed": 2315, "late": 12, "dup": 2665, "predictor": "transition2t70", "m": 8}, "host_time_s": {"sync": 1.3114323328481987, "prefetch": 0.07715808448847383, "ensure": 1.22335886687506, "build": 0.0173785729566589}, "ssd_bytes": 22921740288, "ssd_read_s": 4.247205002582632}24{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2t70", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2t70 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:04:25", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 385.1542919874191, "prompt_tps": 46.77025411947176, "gen_tps": 26.652845588205146, "decode_tps_wall": 26.236517797397315, "itl_ms_p50": 38.85137499310076, "itl_ms_p95": 46.687541995197535, "itl_ms_p99": 51.674208021722734, "itl_ms_mean": 38.114814157967295, "total_s": 2.7863875839393586, "peak_mem_gb": 1.816463, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.5955011046394858, "cache_misses": 4028, "cache_hits": 5930, "stall_s": 1.2055370619054884, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1707, "installed": 1707, "late": 11, "dup": 3369, "predictor": "transition2t70", "m": 8}, "host_time_s": {"sync": 1.3609302886761725, "prefetch": 0.08145119482651353, "ensure": 1.2395877146627754, "build": 0.020705119939520955}, "ssd_bytes": 20295843840, "ssd_read_s": 3.862109358306043}25{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2t50", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2t50 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:04:32", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 518.8054169993848, "prompt_tps": 34.87292168084887, "gen_tps": 29.724905866314426, "decode_tps_wall": 29.261436697346053, "itl_ms_p50": 34.55199999734759, "itl_ms_p95": 41.06304200831801, "itl_ms_p99": 47.81104193534702, "itl_ms_mean": 34.17467195281966, "total_s": 2.6718097500270233, "peak_mem_gb": 1.833023704, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.6617794737899176, "cache_misses": 3368, "cache_hits": 6590, "stall_s": 1.1574277263134718, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3704, "installed": 3698, "late": 153, "dup": 5818, "predictor": "transition2t50", "m": 16}, "host_time_s": {"sync": 1.2977434222120792, "prefetch": 0.13386894366703928, "ensure": 1.1420442478265613, "build": 0.017283538007177413}, "ssd_bytes": 25027411968, "ssd_read_s": 4.566733709187247}26{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2t50", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2t50 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:04:35", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 350.45979102142155, "prompt_tps": 39.98399691941206, "gen_tps": 26.99611487877608, "decode_tps_wall": 26.574413138665893, "itl_ms_p50": 37.45154105126858, "itl_ms_p95": 45.125292032025754, "itl_ms_p99": 48.94620890263468, "itl_ms_mean": 37.630181888946225, "total_s": 2.721161250025034, "peak_mem_gb": 1.77610698, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.6412616339193382, "cache_misses": 3469, "cache_hits": 6201, "stall_s": 1.1320953140966594, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 5175, "installed": 5172, "late": 193, "dup": 5317, "predictor": "transition2t50", "m": 16}, "host_time_s": {"sync": 1.3525510855251923, "prefetch": 0.16570083075203001, "ensure": 1.1026974058477208, "build": 0.017763708136044443}, "ssd_bytes": 30590631936, "ssd_read_s": 4.9587827558862045}27{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2t50", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2t50 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:04:37", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 364.8281251080334, "prompt_tps": 49.38202172368022, "gen_tps": 29.477932572770538, "decode_tps_wall": 29.017480210847932, "itl_ms_p50": 34.84945907257497, "itl_ms_p95": 41.88662499655038, "itl_ms_p99": 45.62962497584522, "itl_ms_mean": 34.46198611091527, "total_s": 2.5359332500956953, "peak_mem_gb": 1.839221404, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.6680056236192007, "cache_misses": 3306, "cache_hits": 6652, "stall_s": 1.0438875223044306, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3642, "installed": 3645, "late": 161, "dup": 5944, "predictor": "transition2t50", "m": 16}, "host_time_s": {"sync": 1.2794099297607318, "prefetch": 0.137731593567878, "ensure": 1.0257551883114502, "build": 0.01715429697651416}, "ssd_bytes": 24588582912, "ssd_read_s": 4.099930818192661}28{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2t50", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2t50 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:04:40", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 341.52458293829113, "prompt_tps": 41.02854651024846, "gen_tps": 27.240125215225166, "decode_tps_wall": 26.81462332201316, "itl_ms_p50": 37.03995898831636, "itl_ms_p95": 44.634749996475875, "itl_ms_p99": 48.89829200692475, "itl_ms_mean": 37.29308400088773, "total_s": 2.6909888749942183, "peak_mem_gb": 1.77610698, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.6412616339193382, "cache_misses": 3469, "cache_hits": 6201, "stall_s": 1.1144671511137858, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 5172, "installed": 5169, "late": 201, "dup": 5320, "predictor": "transition2t50", "m": 16}, "host_time_s": {"sync": 1.34039991220925, "prefetch": 0.16478171071503311, "ensure": 1.0864357101963833, "build": 0.017624183907173574}, "ssd_bytes": 30580015104, "ssd_read_s": 4.9008175508352}29{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2t50", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2t50 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:04:42", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 325.52800001576543, "prompt_tps": 43.04411021589692, "gen_tps": 27.72629489769962, "decode_tps_wall": 27.293189284377878, "itl_ms_p50": 36.34737501852214, "itl_ms_p95": 43.429667013697326, "itl_ms_p99": 47.101125004701316, "itl_ms_mean": 36.63917725336634, "total_s": 2.633796166977845, "peak_mem_gb": 1.78672374, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.6413650465356774, "cache_misses": 3468, "cache_hits": 6202, "stall_s": 1.099088135175407, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 5178, "installed": 5178, "late": 193, "dup": 5314, "predictor": "transition2t50", "m": 16}, "host_time_s": {"sync": 1.3210817894432694, "prefetch": 0.15844549681060016, "ensure": 1.0648446546401829, "build": 0.015979969641193748}, "ssd_bytes": 30597709824, "ssd_read_s": 4.831111097242683}30{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2t50", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor transition2t50 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "d3deca4"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:04:45", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 349.48175004683435, "prompt_tps": 51.5479356880377, "gen_tps": 31.327212216405787, "decode_tps_wall": 30.83788490730665, "itl_ms_p50": 32.85600000526756, "itl_ms_p95": 38.74312504194677, "itl_ms_p99": 43.36958297062665, "itl_ms_mean": 32.427645508303414, "total_s": 2.392423417069949, "peak_mem_gb": 1.835944588, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.6676039365334404, "cache_misses": 3310, "cache_hits": 6648, "stall_s": 0.9897855184972286, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3640, "installed": 3640, "late": 148, "dup": 5946, "predictor": "transition2t50", "m": 16}, "host_time_s": {"sync": 1.2337870801566169, "prefetch": 0.11556095094420016, "ensure": 0.9695895125623792, "build": 0.01276098215021193}, "ssd_bytes": 24595660800, "ssd_read_s": 3.909353203023784}31{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe", "m": 8, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:11:42", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 502.4365830468014, "prompt_tps": 36.03773749574263, "gen_tps": 32.921306889228994, "decode_tps_wall": 32.40808744423553, "itl_ms_p50": 30.22433293517679, "itl_ms_p95": 39.29099999368191, "itl_ms_p99": 41.23612493276596, "itl_ms_mean": 30.85649536464304, "total_s": 2.446395791019313, "peak_mem_gb": 1.84980092, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.7703354087166098, "cache_misses": 2287, "cache_hits": 7671, "stall_s": 0.9387649646960199, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3996, "installed": 3989, "late": 373, "dup": 3924, "predictor": "probe", "m": 8}, "host_time_s": {"sync": 1.2815006269374862, "prefetch": 0.18630608334206045, "ensure": 0.8789908715989441, "build": 0.017593756318092346}, "ssd_bytes": 22235185152, "ssd_read_s": 3.9021120477700606}32{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe", "m": 8, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:11:44", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 373.23191692121327, "prompt_tps": 37.54289365640219, "gen_tps": 31.40608155351881, "decode_tps_wall": 30.915513876587053, "itl_ms_p50": 32.02120796777308, "itl_ms_p95": 39.44387496449053, "itl_ms_p99": 46.47145897615701, "itl_ms_mean": 32.34621957092295, "total_s": 2.411043749889359, "peak_mem_gb": 1.760650756, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.774870734229576, "cache_misses": 2177, "cache_hits": 7493, "stall_s": 0.8995586994569749, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 4130, "installed": 4131, "late": 419, "dup": 4422, "predictor": "probe", "m": 8}, "host_time_s": {"sync": 1.2754781130934134, "prefetch": 0.20737914938945323, "ensure": 0.8304450281430036, "build": 0.01828774716705084}, "ssd_bytes": 22320119808, "ssd_read_s": 3.8726607081480324}33{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe", "m": 8, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:11:47", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 408.2109589362517, "prompt_tps": 44.1305906364464, "gen_tps": 32.024858639512054, "decode_tps_wall": 31.524639137267656, "itl_ms_p50": 31.12475003581494, "itl_ms_p95": 39.56354199908674, "itl_ms_p99": 43.24475000612438, "itl_ms_mean": 31.72121957195775, "total_s": 2.40664779196959, "peak_mem_gb": 1.849231, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.7755573408314923, "cache_misses": 2235, "cache_hits": 7723, "stall_s": 0.9101867240387946, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3949, "installed": 3941, "late": 422, "dup": 4763, "predictor": "probe", "m": 8}, "host_time_s": {"sync": 1.263371414737776, "prefetch": 0.21923305245582014, "ensure": 0.8266534864669666, "build": 0.01808950351551175}, "ssd_bytes": 21884829696, "ssd_read_s": 3.7697257244726643}34{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe", "m": 8, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:11:49", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 356.31954204291105, "prompt_tps": 39.327094646264484, "gen_tps": 31.596870661689568, "decode_tps_wall": 31.10333208308868, "itl_ms_p50": 31.790749984793365, "itl_ms_p95": 38.842375040985644, "itl_ms_p99": 45.3370000468567, "itl_ms_mean": 32.15089615892678, "total_s": 2.381826000055298, "peak_mem_gb": 1.770332132, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.77435367114788, "cache_misses": 2182, "cache_hits": 7488, "stall_s": 0.8886842440115288, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 4122, "installed": 4131, "late": 420, "dup": 4430, "predictor": "probe", "m": 8}, "host_time_s": {"sync": 1.2639832042623311, "prefetch": 0.2039724417263642, "ensure": 0.8186647708062083, "build": 0.017605758039280772}, "ssd_bytes": 22309502976, "ssd_read_s": 3.8125432294327766}35{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe", "m": 8, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:11:52", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 352.57408302277327, "prompt_tps": 39.74218675299524, "gen_tps": 31.65769523405388, "decode_tps_wall": 31.163191468369526, "itl_ms_p50": 32.34412509482354, "itl_ms_p95": 39.54333404544741, "itl_ms_p99": 44.45024998858571, "itl_ms_mean": 32.089139554753075, "total_s": 2.374189874972217, "peak_mem_gb": 1.760503268, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.7731127197518097, "cache_misses": 2194, "cache_hits": 7476, "stall_s": 0.8860083669424057, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 4126, "installed": 4126, "late": 418, "dup": 4426, "predictor": "probe", "m": 8}, "host_time_s": {"sync": 1.2627789897378534, "prefetch": 0.19917341321706772, "ensure": 0.8180779997492209, "build": 0.01713948172982782}, "ssd_bytes": 22366126080, "ssd_read_s": 3.8371908410917968}36{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe", "m": 8, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:11:54", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 412.27854101452976, "prompt_tps": 43.690897816037726, "gen_tps": 32.84528995900421, "decode_tps_wall": 32.33228557038855, "itl_ms_p50": 30.302374972961843, "itl_ms_p95": 39.13604200351983, "itl_ms_p99": 40.90400005225092, "itl_ms_mean": 30.928837301741755, "total_s": 2.3607952910242602, "peak_mem_gb": 1.868302056, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.7755573408314923, "cache_misses": 2235, "cache_hits": 7723, "stall_s": 0.8902485313592479, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3947, "installed": 3940, "late": 439, "dup": 4765, "predictor": "probe", "m": 8}, "host_time_s": {"sync": 1.2474988406756893, "prefetch": 0.20962574891746044, "ensure": 0.8094144120113924, "build": 0.017095185932703316}, "ssd_bytes": 21877751808, "ssd_read_s": 3.686413793475367}37{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe", "m": 16, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:14:14", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 895.8566249348223, "prompt_tps": 20.20077702500302, "gen_tps": 25.055397954692612, "decode_tps_wall": 24.66505483870544, "itl_ms_p50": 38.97029091604054, "itl_ms_p95": 55.66262500360608, "itl_ms_p99": 60.62483391724527, "itl_ms_mean": 40.54318981001242, "total_s": 3.450077582965605, "peak_mem_gb": 1.822079188, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.8403293834103234, "cache_misses": 1590, "cache_hits": 8368, "stall_s": 1.4188859943533316, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 8008, "installed": 7942, "late": 1628, "dup": 7832, "predictor": "probe", "m": 16}, "host_time_s": {"sync": 1.6609962891088799, "prefetch": 0.8776716555003077, "ensure": 0.7751008678460494, "build": 0.022822543629445136}, "ssd_bytes": 33966784512, "ssd_read_s": 6.307455720496364}38{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe", "m": 16, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:14:17", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 443.58262489549816, "prompt_tps": 31.606154157510538, "gen_tps": 24.469822890051635, "decode_tps_wall": 24.087597415190032, "itl_ms_p50": 40.289958007633686, "itl_ms_p95": 52.425541914999485, "itl_ms_p99": 75.69891691673547, "itl_ms_mean": 41.51514087367566, "total_s": 3.059036499937065, "peak_mem_gb": 1.884732324, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.8466390899689762, "cache_misses": 1483, "cache_hits": 8187, "stall_s": 1.1316436724737287, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 8306, "installed": 8290, "late": 1643, "dup": 8166, "predictor": "probe", "m": 16}, "host_time_s": {"sync": 1.553105992032215, "prefetch": 0.8503941696835682, "ensure": 0.5249539959477261, "build": 0.02193720976356417}, "ssd_bytes": 34642722816, "ssd_read_s": 5.387152876704931}39{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe", "m": 16, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:14:20", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 544.7171670384705, "prompt_tps": 33.0735097997283, "gen_tps": 25.842125208025504, "decode_tps_wall": 25.43849308673247, "itl_ms_p50": 37.868833052925766, "itl_ms_p95": 53.40204108506441, "itl_ms_p99": 58.49604192189872, "itl_ms_mean": 39.31050461953476, "total_s": 3.0212789580691606, "peak_mem_gb": 2.003913312, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.8472584856396866, "cache_misses": 1521, "cache_hits": 8437, "stall_s": 1.1443978425813839, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 7934, "installed": 7950, "late": 1592, "dup": 8698, "predictor": "probe", "m": 16}, "host_time_s": {"sync": 1.5227128986734897, "prefetch": 0.826282728696242, "ensure": 0.5458953622728586, "build": 0.021513799438253045}, "ssd_bytes": 33460715520, "ssd_read_s": 5.2902446574298665}40{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe", "m": 16, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:14:23", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 430.2754160016775, "prompt_tps": 32.568850599949556, "gen_tps": 24.76517223527836, "decode_tps_wall": 24.378319392562172, "itl_ms_p50": 40.20533396396786, "itl_ms_p95": 52.76508303359151, "itl_ms_p99": 59.0739999897778, "itl_ms_mean": 41.020054906044926, "total_s": 3.0145388750825077, "peak_mem_gb": 1.88860096, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.8464322647362978, "cache_misses": 1485, "cache_hits": 8185, "stall_s": 1.1297015901654959, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 8301, "installed": 8288, "late": 1629, "dup": 8171, "predictor": "probe", "m": 16}, "host_time_s": {"sync": 1.5254358507227153, "prefetch": 0.8407498337328434, "ensure": 0.5250253080157563, "build": 0.020758838509209454}, "ssd_bytes": 34632105984, "ssd_read_s": 5.336683586472645}41{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe", "m": 16, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:14:26", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 451.5320840291679, "prompt_tps": 31.036623221476617, "gen_tps": 24.792616062079944, "decode_tps_wall": 24.405261377279007, "itl_ms_p50": 41.400167043320835, "itl_ms_p95": 51.83545802719891, "itl_ms_p99": 59.27916697692126, "itl_ms_mean": 40.97477115860711, "total_s": 3.032942667021416, "peak_mem_gb": 1.94522206, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.8471561530506722, "cache_misses": 1478, "cache_hits": 8192, "stall_s": 1.1289817214710638, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 8313, "installed": 8313, "late": 1632, "dup": 8159, "predictor": "probe", "m": 16}, "host_time_s": {"sync": 1.5312410079641268, "prefetch": 0.8357784831896424, "ensure": 0.5366304551716894, "build": 0.02172651037108153}, "ssd_bytes": 34649800704, "ssd_read_s": 5.371664452250116}42{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe", "m": 16, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:14:29", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 504.2627080110833, "prompt_tps": 35.728608423431574, "gen_tps": 26.203663395431743, "decode_tps_wall": 25.79470949171051, "itl_ms_p50": 37.67937491647899, "itl_ms_p95": 53.558959043584764, "itl_ms_p99": 58.460749918594956, "itl_ms_mean": 38.76763955497789, "total_s": 2.9466239999746904, "peak_mem_gb": 1.9931818, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.8467563767824865, "cache_misses": 1526, "cache_hits": 8432, "stall_s": 1.1152964674402028, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 7921, "installed": 7932, "late": 1617, "dup": 8711, "predictor": "probe", "m": 16}, "host_time_s": {"sync": 1.487131759757176, "prefetch": 0.8097676509059966, "ensure": 0.5279155237367377, "build": 0.020733201992698014}, "ssd_bytes": 33432403968, "ssd_read_s": 5.181995438179001}43{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe2", "m": 8, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:17:17", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 1001.3276670360938, "prompt_tps": 18.062219831373454, "gen_tps": 30.189756179784066, "decode_tps_wall": 29.71969025036037, "itl_ms_p50": 32.692833920009434, "itl_ms_p95": 43.190291966311634, "itl_ms_p99": 56.555208982899785, "itl_ms_mean": 33.647726190143395, "total_s": 3.121134417015128, "peak_mem_gb": 1.846261988, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.7363928499698734, "cache_misses": 2625, "cache_hits": 7333, "stall_s": 1.3216708386316895, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3701, "installed": 3695, "late": 361, "dup": 4947, "predictor": "probe2", "m": 8}, "host_time_s": {"sync": 1.5726819206029177, "prefetch": 0.21285899879876524, "ensure": 1.2344930110266432, "build": 0.018278107163496315}, "ssd_bytes": 22387359744, "ssd_read_s": 5.210404163459316}44{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe2", "m": 8, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:17:20", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 399.63033294770867, "prompt_tps": 35.083029868527234, "gen_tps": 30.869702613755237, "decode_tps_wall": 30.387516808679806, "itl_ms_p50": 33.09150005225092, "itl_ms_p95": 39.47520803194493, "itl_ms_p99": 40.891999960877, "itl_ms_mean": 32.90824999936694, "total_s": 2.4728500829078257, "peak_mem_gb": 1.76247242, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.7378490175801448, "cache_misses": 2535, "cache_hits": 7135, "stall_s": 0.9910381255904213, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3868, "installed": 3869, "late": 344, "dup": 4684, "predictor": "probe2", "m": 8}, "host_time_s": {"sync": 1.2793637560680509, "prefetch": 0.18615518277511, "ensure": 0.9213677381630987, "build": 0.01609112205915153}, "ssd_bytes": 22659858432, "ssd_read_s": 4.111017375951633}45{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe2", "m": 8, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:17:22", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 427.46583302505314, "prompt_tps": 42.14546568691707, "gen_tps": 31.554680624149604, "decode_tps_wall": 31.061781672296963, "itl_ms_p50": 32.138042035512626, "itl_ms_p95": 39.94012507610023, "itl_ms_p99": 45.22000008728355, "itl_ms_mean": 32.193903445399236, "total_s": 2.455681750085205, "peak_mem_gb": 1.830553232, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.7414139385418759, "cache_misses": 2575, "cache_hits": 7383, "stall_s": 1.0046712778275833, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3647, "installed": 3644, "late": 350, "dup": 5065, "predictor": "probe2", "m": 8}, "host_time_s": {"sync": 1.253214834840037, "prefetch": 0.1883800079813227, "ensure": 0.9295253350865096, "build": 0.015667336992919445}, "ssd_bytes": 22019309568, "ssd_read_s": 4.02383333875332}46{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe2", "m": 8, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:17:24", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 396.5723329456523, "prompt_tps": 35.327768118162375, "gen_tps": 30.97991967772089, "decode_tps_wall": 30.49599742967766, "itl_ms_p50": 32.31929091271013, "itl_ms_p95": 41.01816599722952, "itl_ms_p99": 45.103374985046685, "itl_ms_mean": 32.79118849304579, "total_s": 2.462417208007537, "peak_mem_gb": 1.755394532, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.7381592554291624, "cache_misses": 2532, "cache_hits": 7138, "stall_s": 0.9832507125101984, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3866, "installed": 3869, "late": 358, "dup": 4686, "predictor": "probe2", "m": 8}, "host_time_s": {"sync": 1.2777193939546123, "prefetch": 0.18355376191902906, "ensure": 0.916660810704343, "build": 0.015516375424340367}, "ssd_bytes": 22642163712, "ssd_read_s": 4.108548927935772}47{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe2", "m": 8, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:17:27", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 401.68699994683266, "prompt_tps": 34.87923066595009, "gen_tps": 31.12611108532413, "decode_tps_wall": 30.639938206164455, "itl_ms_p50": 31.847332953475416, "itl_ms_p95": 40.61833396553993, "itl_ms_p99": 45.17837497405708, "itl_ms_mean": 32.63714153962653, "total_s": 2.4578269169433042, "peak_mem_gb": 1.762732952, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.7381592554291624, "cache_misses": 2532, "cache_hits": 7138, "stall_s": 0.9817315798718482, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3873, "installed": 3874, "late": 347, "dup": 4679, "predictor": "probe2", "m": 8}, "host_time_s": {"sync": 1.2711185896769166, "prefetch": 0.18561373895499855, "ensure": 0.9145922709722072, "build": 0.016176007338799536}, "ssd_bytes": 22666936320, "ssd_read_s": 4.091072851326317}48{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe2", "m": 8, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:17:29", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 466.777375084348, "prompt_tps": 38.59071709355048, "gen_tps": 32.02471239861806, "decode_tps_wall": 31.524491235271036, "itl_ms_p50": 31.066416995599866, "itl_ms_p95": 38.839290966279805, "itl_ms_p99": 43.48008299712092, "itl_ms_mean": 31.7213683969356, "total_s": 2.465223584091291, "peak_mem_gb": 1.828194016, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.741715203856196, "cache_misses": 2572, "cache_hits": 7386, "stall_s": 0.9836312148254365, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3648, "installed": 3644, "late": 356, "dup": 5064, "predictor": "probe2", "m": 8}, "host_time_s": {"sync": 1.288077262463048, "prefetch": 0.18259123258758336, "ensure": 0.9126184373162687, "build": 0.01517882663756609}, "ssd_bytes": 22012231680, "ssd_read_s": 3.9626905703917146}49{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe3", "m": 8, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:19:01", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 1018.3106670156121, "prompt_tps": 17.776447098219712, "gen_tps": 31.891249500646396, "decode_tps_wall": 31.3950959073452, "itl_ms_p50": 30.84979101549834, "itl_ms_p95": 40.79133295454085, "itl_ms_p99": 43.997166911140084, "itl_ms_mean": 31.852108461501462, "total_s": 3.024993500090204, "peak_mem_gb": 1.8370214, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.7214300060253063, "cache_misses": 2774, "cache_hits": 7184, "stall_s": 1.2297115926630795, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3490, "installed": 3482, "late": 341, "dup": 5086, "predictor": "probe3", "m": 8}, "host_time_s": {"sync": 1.5910293539054692, "prefetch": 0.1771570744458586, "ensure": 1.1629536031978205, "build": 0.015928100096061826}, "ssd_bytes": 22167945216, "ssd_read_s": 4.735749578336254}50{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe3", "m": 8, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:19:03", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 417.7632079226896, "prompt_tps": 33.55570644273465, "gen_tps": 30.655205930220156, "decode_tps_wall": 30.176362270562443, "itl_ms_p50": 33.24174997396767, "itl_ms_p95": 40.57654202915728, "itl_ms_p99": 42.76920796837658, "itl_ms_mean": 33.138520509329815, "total_s": 2.505490000010468, "peak_mem_gb": 1.76902902, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.7127197518097208, "cache_misses": 2778, "cache_hits": 6892, "stall_s": 1.019743283977732, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3569, "installed": 3575, "late": 290, "dup": 4983, "predictor": "probe3", "m": 8}, "host_time_s": {"sync": 1.2866521296091378, "prefetch": 0.16383414575830102, "ensure": 0.9688517367467284, "build": 0.015919991652481258}, "ssd_bytes": 22461677568, "ssd_read_s": 4.080189808388241}51{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe3", "m": 8, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:19:06", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 513.5413750540465, "prompt_tps": 35.07477991648297, "gen_tps": 31.377521603349404, "decode_tps_wall": 30.88737086696359, "itl_ms_p50": 31.350291916169226, "itl_ms_p95": 40.156333008781075, "itl_ms_p99": 47.554834047332406, "itl_ms_mean": 32.3756918096767, "total_s": 2.553209959063679, "peak_mem_gb": 1.821116128, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.7267523599116288, "cache_misses": 2721, "cache_hits": 7237, "stall_s": 1.0283256021793932, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3440, "installed": 3432, "late": 342, "dup": 5272, "predictor": "probe3", "m": 8}, "host_time_s": {"sync": 1.3280688052764162, "prefetch": 0.17529725469648838, "ensure": 0.9640046572312713, "build": 0.015673740301281214}, "ssd_bytes": 21803433984, "ssd_read_s": 4.0124820349738}52{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe3", "m": 8, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:19:08", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 445.1344159897417, "prompt_tps": 31.47338430791281, "gen_tps": 29.277490438874107, "decode_tps_wall": 28.820232896597577, "itl_ms_p50": 34.61512504145503, "itl_ms_p95": 43.48645801655948, "itl_ms_p99": 49.46470900904387, "itl_ms_mean": 34.69784590526528, "total_s": 2.6310987080214545, "peak_mem_gb": 1.76902902, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.7127197518097208, "cache_misses": 2778, "cache_hits": 6892, "stall_s": 1.0517092497320846, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3566, "installed": 3574, "late": 288, "dup": 4986, "predictor": "probe3", "m": 8}, "host_time_s": {"sync": 1.36907018115744, "prefetch": 0.1682968153618276, "ensure": 1.001935888431035, "build": 0.017286069924011827}, "ssd_bytes": 22451060736, "ssd_read_s": 4.230282494449057}53{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe3", "m": 8, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:19:11", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 422.8654580656439, "prompt_tps": 33.132170410624965, "gen_tps": 30.11652802794099, "decode_tps_wall": 29.646075272016446, "itl_ms_p50": 33.69879210367799, "itl_ms_p95": 42.30249999091029, "itl_ms_p99": 46.123499982059, "itl_ms_mean": 33.73127777705945, "total_s": 2.547935958020389, "peak_mem_gb": 1.76177274, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.7116856256463289, "cache_misses": 2788, "cache_hits": 6882, "stall_s": 1.032603504951112, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3573, "installed": 3572, "late": 301, "dup": 4979, "predictor": "probe3", "m": 8}, "host_time_s": {"sync": 1.3089445441728458, "prefetch": 0.1695250024786219, "ensure": 0.9806344367098063, "build": 0.01697311399038881}, "ssd_bytes": 22511222784, "ssd_read_s": 4.126012323773466}54{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "probe3", "m": 8, "prefetch_workers": 2, "fused_upgate": false, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "attribution", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor probe3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general --split heldout --repeats 2 --max-tokens 64 --tag attribution --out benchmarks/raw/gateC/cold_nvme_laptop/attribution.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "7b5afc7"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:19:13", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 487.9956250078976, "prompt_tps": 36.9404389279317, "gen_tps": 32.56493375331503, "decode_tps_wall": 32.05627861299624, "itl_ms_p50": 30.27679189108312, "itl_ms_p95": 37.66083298251033, "itl_ms_p99": 42.33550000935793, "itl_ms_mean": 31.195136905086063, "total_s": 2.4532892500283197, "peak_mem_gb": 1.820984976, "text": "\n\nMac Studio:\n- More powerful hardware options, including multiple GPUs and more RAM.\n- Higher-end CPU options, often with more cores and higher clock speeds.\n- Larger display with a higher resolution", "cache_hit_rate": 0.7265515163687487, "cache_misses": 2723, "cache_hits": 7235, "stall_s": 1.0000087718944997, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3441, "installed": 3438, "late": 337, "dup": 5271, "predictor": "probe3", "m": 8}, "host_time_s": {"sync": 1.2687757798703387, "prefetch": 0.17196745437104255, "ensure": 0.9326372838113457, "build": 0.014673922210931778}, "ssd_bytes": 21814050816, "ssd_read_s": 3.9007015310926363}55