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": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "8a2444c"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:50:48", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 692.0957081019878, "prompt_tps": 46.59402246111043, "gen_tps": 36.57738284532069, "decode_tps_wall": 36.007635327202976, "itl_ms_p50": 27.320333058014512, "itl_ms_p95": 33.130957977846265, "itl_ms_p99": 39.143749978393316, "itl_ms_mean": 27.77188757087089, "total_s": 2.441724625066854, "peak_mem_gb": 2.090173904, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.4844649021864212, "cache_misses": 5824, "cache_hits": 5473, "stall_s": 1.3842638342175633, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 20610809856, "ssd_read_s": 3.7224394695367664}2{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "8a2444c"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:50:50", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 369.74300001747906, "prompt_tps": 64.96787686611094, "gen_tps": 44.10237031807938, "decode_tps_wall": 43.4135612400075, "itl_ms_p50": 22.59599999524653, "itl_ms_p95": 29.346165945753455, "itl_ms_p99": 34.49933393858373, "itl_ms_mean": 23.034277111513628, "total_s": 1.8209024580428377, "peak_mem_gb": 1.9139122, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.6523887973640856, "cache_misses": 3587, "cache_hits": 6732, "stall_s": 0.9014140029903501, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 12694192128, "ssd_read_s": 2.055849175201729}3{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "8a2444c"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:50:52", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 316.1524999886751, "prompt_tps": 44.3261007825923, "gen_tps": 37.43535274314916, "decode_tps_wall": 36.85073790399236, "itl_ms_p50": 26.99137502349913, "itl_ms_p95": 33.23575004469603, "itl_ms_p99": 36.866707960143685, "itl_ms_mean": 27.136498666737996, "total_s": 2.025751915993169, "peak_mem_gb": 1.737943192, "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": 0.48355739400206826, "cache_misses": 4994, "cache_hits": 4676, "stall_s": 1.0538058636011556, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 17673486336, "ssd_read_s": 2.6701927126850933}4{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "8a2444c"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:50:55", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 345.18750000279397, "prompt_tps": 52.19889665094133, "gen_tps": 35.34670805642188, "decode_tps_wall": 34.79462232635829, "itl_ms_p50": 28.336417046375573, "itl_ms_p95": 35.60750000178814, "itl_ms_p99": 37.13658300694078, "itl_ms_mean": 28.74007341193242, "total_s": 2.1558121249545366, "peak_mem_gb": 1.814131732, "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": 1.1506640871521086, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 17439916032, "ssd_read_s": 2.9321806946536526}5{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "8a2444c"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:50:57", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 420.1389999361709, "prompt_tps": 69.07719609252912, "gen_tps": 37.35810759487954, "decode_tps_wall": 36.77462865794749, "itl_ms_p50": 28.77229091245681, "itl_ms_p95": 33.97729201242328, "itl_ms_p99": 35.62474995851517, "itl_ms_mean": 27.192660714573567, "total_s": 2.1332766249543056, "peak_mem_gb": 1.905393932, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.5451856093138658, "cache_misses": 4766, "cache_hits": 5713, "stall_s": 1.1352890465641394, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 16866607104, "ssd_read_s": 2.85799129109364}6{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "8a2444c"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:50:58", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 325.22700005210936, "prompt_tps": 58.47830967620645, "gen_tps": 44.22672637332204, "decode_tps_wall": 43.53605355216225, "itl_ms_p50": 21.73804200720042, "itl_ms_p95": 29.404125059954822, "itl_ms_p99": 31.615916988812387, "itl_ms_mean": 22.96946825466991, "total_s": 1.7723035000963137, "peak_mem_gb": 1.773794588, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.6699147381242387, "cache_misses": 3252, "cache_hits": 6600, "stall_s": 0.8593332751188427, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 11508645888, "ssd_read_s": 1.9043039631797}7{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "8a2444c"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:51:01", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 958.3374169887975, "prompt_tps": 33.402911106697594, "gen_tps": 33.07673168156339, "decode_tps_wall": 32.560083043862754, "itl_ms_p50": 29.210791923105717, "itl_ms_p95": 35.54279205854982, "itl_ms_p99": 116.7389169568196, "itl_ms_mean": 30.712452380814486, "total_s": 2.89322191698011, "peak_mem_gb": 2.094958152, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.49030716119323714, "cache_misses": 5758, "cache_hits": 5539, "stall_s": 1.7997499047778547, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 20377239552, "ssd_read_s": 4.840675379149616}8{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "8a2444c"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:51:03", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 368.4726250357926, "prompt_tps": 48.89591995154156, "gen_tps": 35.25698277643246, "decode_tps_wall": 34.70629636717931, "itl_ms_p50": 28.44312507659197, "itl_ms_p95": 35.80216702539474, "itl_ms_p99": 36.164957913570106, "itl_ms_mean": 28.81321560273627, "total_s": 2.1837052080081776, "peak_mem_gb": 1.839953, "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": 0.5069291022293633, "cache_misses": 4910, "cache_hits": 5048, "stall_s": 1.1720648780465126, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 17376215040, "ssd_read_s": 2.986359332106076}9{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "8a2444c"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:51:06", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 335.31462494283915, "prompt_tps": 41.79292133374972, "gen_tps": 34.24103356449932, "decode_tps_wall": 33.70621428467988, "itl_ms_p50": 29.802417033351958, "itl_ms_p95": 36.68737504631281, "itl_ms_p99": 40.050665964372456, "itl_ms_mean": 29.66811969905855, "total_s": 2.204406165983528, "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": 1.1822265810333192, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 17811505152, "ssd_read_s": 3.05972489528358}10{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "8a2444c"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:51:08", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 418.8025000039488, "prompt_tps": 69.29881355560856, "gen_tps": 37.537740267744965, "decode_tps_wall": 36.95148489956471, "itl_ms_p50": 28.28020800370723, "itl_ms_p95": 33.74383400660008, "itl_ms_p99": 34.99762492720038, "itl_ms_mean": 27.062511904948643, "total_s": 2.1237407500157133, "peak_mem_gb": 1.905393956, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.5457581830327322, "cache_misses": 4760, "cache_hits": 5719, "stall_s": 1.127805772703141, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 16845373440, "ssd_read_s": 2.838174633914605}11{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "8a2444c"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:51:10", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 374.44504199083894, "prompt_tps": 64.15261188613987, "gen_tps": 42.52224456988057, "decode_tps_wall": 41.858079039022705, "itl_ms_p50": 24.035749956965446, "itl_ms_p95": 30.22029099520296, "itl_ms_p99": 36.48670797701925, "itl_ms_mean": 23.890250650722358, "total_s": 1.8795308329863474, "peak_mem_gb": 1.925774224, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.6497722647543367, "cache_misses": 3614, "cache_hits": 6705, "stall_s": 0.9495816181879491, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 12789743616, "ssd_read_s": 2.1544746939325705}12{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "8a2444c"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:51:11", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 316.6579579701647, "prompt_tps": 60.06138905385063, "gen_tps": 43.74587372657144, "decode_tps_wall": 43.062635113920706, "itl_ms_p50": 21.923959022387862, "itl_ms_p95": 30.937374918721616, "itl_ms_p99": 31.97908296715468, "itl_ms_mean": 23.221988096049735, "total_s": 1.779643208021298, "peak_mem_gb": 1.783100776, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.6759033698741372, "cache_misses": 3193, "cache_hits": 6659, "stall_s": 0.8607746257912368, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 11299848192, "ssd_read_s": 1.8981057128403336}13{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "8a2444c"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:51:14", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 416.65312508121133, "prompt_tps": 69.65812977727899, "gen_tps": 37.280073973973195, "decode_tps_wall": 36.69780081939727, "itl_ms_p50": 28.502625063993037, "itl_ms_p95": 34.91549997124821, "itl_ms_p99": 36.375541938468814, "itl_ms_mean": 27.249589285236738, "total_s": 2.133377250051126, "peak_mem_gb": 1.903411428, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.5446130355949995, "cache_misses": 4772, "cache_hits": 5707, "stall_s": 1.1383883256930858, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 16887840768, "ssd_read_s": 2.8613588492153212}14{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "8a2444c"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:51:15", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 341.64862509351224, "prompt_tps": 55.65981600473538, "gen_tps": 43.30100596604308, "decode_tps_wall": 42.624710140036576, "itl_ms_p50": 22.45845808647573, "itl_ms_p95": 29.735749936662614, "itl_ms_p99": 32.036874908953905, "itl_ms_mean": 23.460570094545208, "total_s": 1.8196645410498604, "peak_mem_gb": 1.773794588, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.6699147381242387, "cache_misses": 3252, "cache_hits": 6600, "stall_s": 0.8717616148060188, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 11508645888, "ssd_read_s": 1.923264330253005}15{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "8a2444c"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:51:18", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 362.4877498950809, "prompt_tps": 49.698369966583876, "gen_tps": 35.44949672632509, "decode_tps_wall": 34.895773922080814, "itl_ms_p50": 28.025666950270534, "itl_ms_p95": 36.12537495791912, "itl_ms_p99": 37.00979193672538, "itl_ms_mean": 28.65676520695348, "total_s": 2.1678639579331502, "peak_mem_gb": 1.843491944, "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": 0.5079333199437638, "cache_misses": 4900, "cache_hits": 5058, "stall_s": 1.1607553774956614, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 17340825600, "ssd_read_s": 2.963632360799238}16{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "8a2444c"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:51:20", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 401.48825000505894, "prompt_tps": 59.82431586713207, "gen_tps": 41.56776955586064, "decode_tps_wall": 40.91847359488166, "itl_ms_p50": 23.810542072169483, "itl_ms_p95": 29.90987489465624, "itl_ms_p99": 36.111374967731535, "itl_ms_mean": 24.438839285664024, "total_s": 1.9411351250018924, "peak_mem_gb": 1.934097296, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.6497722647543367, "cache_misses": 3614, "cache_hits": 6705, "stall_s": 0.9638368771411479, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 12789743616, "ssd_read_s": 2.177788548753597}17{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "8a2444c"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:51:22", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 334.01695801876485, "prompt_tps": 41.95431694931566, "gen_tps": 33.55691134043318, "decode_tps_wall": 33.0327910023831, "itl_ms_p50": 30.482040951028466, "itl_ms_p95": 37.04166691750288, "itl_ms_p99": 41.51424998417497, "itl_ms_mean": 30.272949080441204, "total_s": 2.241212750086561, "peak_mem_gb": 1.742661784, "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": 0.48355739400206826, "cache_misses": 4994, "cache_hits": 4676, "stall_s": 1.1990491144824773, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 17673486336, "ssd_read_s": 3.0536379632540047}18{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "8a2444c"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:51:24", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 575.3412500489503, "prompt_tps": 55.65209726540678, "gen_tps": 34.034027323611205, "decode_tps_wall": 33.502500274535386, "itl_ms_p50": 29.327333089895546, "itl_ms_p95": 36.18929197546095, "itl_ms_p99": 44.5929579436779, "itl_ms_mean": 29.848518522663248, "total_s": 2.455797916976735, "peak_mem_gb": 2.079360592, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.49048419934495885, "cache_misses": 5756, "cache_hits": 5541, "stall_s": 1.38695521466434, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 20370161664, "ssd_read_s": 3.5395298274233937}19{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:51:33", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 772.1267499728128, "prompt_tps": 41.608416670316196, "gen_tps": 28.47543362667651, "decode_tps_wall": 28.03176881410257, "itl_ms_p50": 34.94879195932299, "itl_ms_p95": 42.67483309376985, "itl_ms_p99": 57.45962506625801, "itl_ms_mean": 35.67381019127511, "total_s": 3.0195767920231447, "peak_mem_gb": 2.160502028, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.6128175621846508, "cache_misses": 4374, "cache_hits": 6923, "stall_s": 1.482811972266063, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3581, "installed": 3580, "late": 46, "dup": 5803, "predictor": "transition", "m": 8}, "ssd_bytes": 28152299520, "ssd_read_s": 5.722053766134195}20{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:51:35", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 408.1629170104861, "prompt_tps": 58.852348604539685, "gen_tps": 35.390164693827245, "decode_tps_wall": 34.83739886083086, "itl_ms_p50": 28.134417021647096, "itl_ms_p95": 37.53758396487683, "itl_ms_p99": 40.30858399346471, "itl_ms_mean": 28.704783729543642, "total_s": 2.2165642919717357, "peak_mem_gb": 1.9614018, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.7241011725942437, "cache_misses": 2847, "cache_hits": 7472, "stall_s": 0.9345974458847195, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1557, "installed": 1558, "late": 20, "dup": 7339, "predictor": "transition", "m": 8}, "ssd_bytes": 15585509376, "ssd_read_s": 2.9250047012465075}21{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:51:38", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 378.2986250007525, "prompt_tps": 37.04016045882731, "gen_tps": 27.031827230294134, "decode_tps_wall": 26.609577156162334, "itl_ms_p50": 37.20854106359184, "itl_ms_p95": 44.75649993401021, "itl_ms_p99": 47.99770808313042, "itl_ms_mean": 37.58045436540944, "total_s": 2.745867250021547, "peak_mem_gb": 1.789864856, "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": 0.6477766287487073, "cache_misses": 3406, "cache_hits": 6264, "stall_s": 1.2254173471592367, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 4572, "installed": 4565, "late": 112, "dup": 3980, "predictor": "transition", "m": 8}, "ssd_bytes": 28233695232, "ssd_read_s": 5.215782183106057}22{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:51:40", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 395.0674580410123, "prompt_tps": 45.60141609241971, "gen_tps": 28.743164707500142, "decode_tps_wall": 28.294193060592328, "itl_ms_p50": 35.50791600719094, "itl_ms_p95": 42.84283402375877, "itl_ms_p99": 45.48420803621411, "itl_ms_mean": 35.34294114196821, "total_s": 2.6216727499850094, "peak_mem_gb": 1.89117404, "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": 0.6738300863627235, "cache_misses": 3248, "cache_hits": 6710, "stall_s": 1.164204154862091, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3562, "installed": 3568, "late": 123, "dup": 5150, "predictor": "transition", "m": 8}, "ssd_bytes": 24100208640, "ssd_read_s": 4.4939437442226335}23{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:51:43", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 459.74879094865173, "prompt_tps": 63.12747100100931, "gen_tps": 30.265578685281362, "decode_tps_wall": 29.79284338060757, "itl_ms_p50": 34.34629202820361, "itl_ms_p95": 42.33533306978643, "itl_ms_p99": 44.85229204874486, "itl_ms_mean": 33.5651078087736, "total_s": 2.5743505829013884, "peak_mem_gb": 1.962727656, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.6650443744632122, "cache_misses": 3510, "cache_hits": 6969, "stall_s": 1.1666903716977686, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2836, "installed": 2837, "late": 39, "dup": 6084, "predictor": "transition", "m": 8}, "ssd_bytes": 22458138624, "ssd_read_s": 4.246741865412332}24{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:51:45", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 365.00933300703764, "prompt_tps": 52.09831087984164, "gen_tps": 34.83227387710879, "decode_tps_wall": 34.28816965332164, "itl_ms_p50": 27.327749994583428, "itl_ms_p95": 37.82312502153218, "itl_ms_p99": 43.85545803233981, "itl_ms_mean": 29.164578048660168, "total_s": 2.2023777500726283, "peak_mem_gb": 1.811433372, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.749187982135607, "cache_misses": 2471, "cache_hits": 7381, "stall_s": 0.8500924549298361, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1530, "installed": 1530, "late": 15, "dup": 7134, "predictor": "transition", "m": 8}, "ssd_bytes": 14159314944, "ssd_read_s": 2.5876708430005237}25{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:51:48", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 609.2141249682754, "prompt_tps": 52.55432495384122, "gen_tps": 27.046209505045038, "decode_tps_wall": 26.623845944181046, "itl_ms_p50": 37.319499999284744, "itl_ms_p95": 44.7193750878796, "itl_ms_p99": 57.728291023522615, "itl_ms_mean": 37.56031349101769, "total_s": 2.97551387490239, "peak_mem_gb": 2.134008904, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.6143223864742852, "cache_misses": 4357, "cache_hits": 6940, "stall_s": 1.4290780789451674, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3428, "installed": 3427, "late": 59, "dup": 5956, "predictor": "transition", "m": 8}, "ssd_bytes": 27550679040, "ssd_read_s": 5.290976330987178}26{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:51:51", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 416.3769589504227, "prompt_tps": 43.264303739765175, "gen_tps": 27.988423868178053, "decode_tps_wall": 27.55124934041564, "itl_ms_p50": 36.89020802266896, "itl_ms_p95": 44.10533409100026, "itl_ms_p99": 46.77029198501259, "itl_ms_mean": 36.29599469861696, "total_s": 2.703024624963291, "peak_mem_gb": 1.896056532, "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": 0.6744326169913637, "cache_misses": 3242, "cache_hits": 6716, "stall_s": 1.2099400870501995, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3566, "installed": 3566, "late": 130, "dup": 5146, "predictor": "transition", "m": 8}, "ssd_bytes": 24093130752, "ssd_read_s": 4.651791002950631}27{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:51:54", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 381.9474159972742, "prompt_tps": 36.684246394513124, "gen_tps": 25.94018958681683, "decode_tps_wall": 25.53496770643889, "itl_ms_p50": 38.30345894675702, "itl_ms_p95": 46.70962505042553, "itl_ms_p99": 50.887249992229044, "itl_ms_mean": 39.161984126881826, "total_s": 2.8491524159908295, "peak_mem_gb": 1.79723612, "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": 0.6458117890382626, "cache_misses": 3425, "cache_hits": 6245, "stall_s": 1.2707051890902221, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 4557, "installed": 4550, "late": 119, "dup": 3995, "predictor": "transition", "m": 8}, "ssd_bytes": 28247851008, "ssd_read_s": 5.366749788634479}28{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:51:56", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 481.4576250500977, "prompt_tps": 60.28089337608887, "gen_tps": 28.671537883760458, "decode_tps_wall": 28.223687879672383, "itl_ms_p50": 35.76179197989404, "itl_ms_p95": 44.94012508075684, "itl_ms_p99": 48.35233394987881, "itl_ms_mean": 35.43123082509116, "total_s": 2.713625167030841, "peak_mem_gb": 1.956036108, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.6653306613226453, "cache_misses": 3507, "cache_hits": 6972, "stall_s": 1.233517144457437, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2849, "installed": 2857, "late": 41, "dup": 6071, "predictor": "transition", "m": 8}, "ssd_bytes": 22493528064, "ssd_read_s": 4.465901253512129}29{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:51:59", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 442.5310830119997, "prompt_tps": 54.27066306043648, "gen_tps": 33.787125666716776, "decode_tps_wall": 33.25937302301062, "itl_ms_p50": 29.701624996960163, "itl_ms_p95": 38.44520903658122, "itl_ms_p99": 39.874125039204955, "itl_ms_mean": 30.06671230116534, "total_s": 2.336733957985416, "peak_mem_gb": 1.94344492, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.7230351778273089, "cache_misses": 2858, "cache_hits": 7461, "stall_s": 0.9875611962052062, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1567, "installed": 1567, "late": 19, "dup": 7329, "predictor": "transition", "m": 8}, "ssd_bytes": 15659827200, "ssd_read_s": 3.0826816788176075}30{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:52:01", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 367.9271249566227, "prompt_tps": 51.68762347875283, "gen_tps": 36.96160404975749, "decode_tps_wall": 36.384369678357146, "itl_ms_p50": 26.697875000536442, "itl_ms_p95": 36.270500044338405, "itl_ms_p99": 39.05266697984189, "itl_ms_mean": 27.484329365607763, "total_s": 2.099439874989912, "peak_mem_gb": 1.8149887, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.750812017864393, "cache_misses": 2455, "cache_hits": 7397, "stall_s": 0.8192729116417468, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1505, "installed": 1505, "late": 20, "dup": 7159, "predictor": "transition", "m": 8}, "ssd_bytes": 14014218240, "ssd_read_s": 2.4920683407690376}31{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:52:04", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 453.00583308562636, "prompt_tps": 64.06463609999683, "gen_tps": 26.82671878297458, "decode_tps_wall": 26.407680900724202, "itl_ms_p50": 35.80704203341156, "itl_ms_p95": 63.98283294402063, "itl_ms_p99": 102.152582956478, "itl_ms_mean": 37.8677705081091, "total_s": 2.8386753750965, "peak_mem_gb": 1.947228324, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.6648535165569234, "cache_misses": 3512, "cache_hits": 6967, "stall_s": 1.2396353124640882, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2843, "installed": 2843, "late": 44, "dup": 6077, "predictor": "transition", "m": 8}, "ssd_bytes": 22489989120, "ssd_read_s": 4.380332345608622}32{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:52:06", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 431.1944170622155, "prompt_tps": 44.09987648009074, "gen_tps": 28.18513959548602, "decode_tps_wall": 27.744889843902232, "itl_ms_p50": 35.49208294134587, "itl_ms_p95": 45.85787502583116, "itl_ms_p99": 49.1723750019446, "itl_ms_mean": 36.042673286006206, "total_s": 2.7018828340806067, "peak_mem_gb": 1.809139612, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.7490864799025578, "cache_misses": 2472, "cache_hits": 7380, "stall_s": 1.046185417799279, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1530, "installed": 1530, "late": 20, "dup": 7134, "predictor": "transition", "m": 8}, "ssd_bytes": 14162853888, "ssd_read_s": 3.140761493239552}33{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:52:09", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 450.0240000197664, "prompt_tps": 40.031146461605665, "gen_tps": 26.194949735952612, "decode_tps_wall": 25.785786902003856, "itl_ms_p50": 37.66541602090001, "itl_ms_p95": 47.37641604151577, "itl_ms_p99": 52.02454107347876, "itl_ms_mean": 38.78105422186236, "total_s": 2.8932304159970954, "peak_mem_gb": 1.86305908, "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": 0.6752359911628841, "cache_misses": 3234, "cache_hits": 6724, "stall_s": 1.2523563622962683, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3532, "installed": 3531, "late": 126, "dup": 5180, "predictor": "transition", "m": 8}, "ssd_bytes": 23944495104, "ssd_read_s": 4.7414319039089605}34{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:52:12", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 449.3287919322029, "prompt_tps": 53.458818051430605, "gen_tps": 32.35508969994482, "decode_tps_wall": 31.849744706017475, "itl_ms_p50": 30.898042023181915, "itl_ms_p95": 41.62845900282264, "itl_ms_p99": 43.35575003642589, "itl_ms_mean": 31.397425920687734, "total_s": 2.42736662493553, "peak_mem_gb": 1.98237332, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.7226475433666053, "cache_misses": 2862, "cache_hits": 7457, "stall_s": 0.9958835493307561, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1564, "installed": 1565, "late": 22, "dup": 7332, "predictor": "transition", "m": 8}, "ssd_bytes": 15663366144, "ssd_read_s": 3.0712371243862435}35{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:52:15", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 391.460333019495, "prompt_tps": 35.791297177265804, "gen_tps": 25.000425221607767, "decode_tps_wall": 24.609897719801413, "itl_ms_p50": 40.51454202271998, "itl_ms_p95": 48.93241706304252, "itl_ms_p99": 52.02154198195785, "itl_ms_mean": 40.634057540003035, "total_s": 2.951405958039686, "peak_mem_gb": 1.789864856, "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": 0.6473629782833505, "cache_misses": 3410, "cache_hits": 6260, "stall_s": 1.2819557231850922, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 4557, "installed": 4552, "late": 119, "dup": 3995, "predictor": "transition", "m": 8}, "ssd_bytes": 28194766848, "ssd_read_s": 5.415862058289349}36{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:52:18", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 619.5360409328714, "prompt_tps": 51.67940227960384, "gen_tps": 27.820224544948196, "decode_tps_wall": 27.385726982564265, "itl_ms_p50": 35.614707972854376, "itl_ms_p95": 45.39079195819795, "itl_ms_p99": 54.132417077198625, "itl_ms_mean": 36.515371698427884, "total_s": 2.9200044579338282, "peak_mem_gb": 2.13448616, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.6143223864742852, "cache_misses": 4357, "cache_hits": 6940, "stall_s": 1.4040574849350378, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3430, "installed": 3434, "late": 63, "dup": 5954, "predictor": "transition", "m": 8}, "ssd_bytes": 27557756928, "ssd_read_s": 5.245501340832561}37{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:52:26", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 771.1540409363806, "prompt_tps": 41.65372223002491, "gen_tps": 27.094916876170924, "decode_tps_wall": 26.672521919988778, "itl_ms_p50": 37.9937500692904, "itl_ms_p95": 46.87191604170948, "itl_ms_p99": 64.66595910023898, "itl_ms_mean": 37.49176785756376, "total_s": 3.1331354159628972, "peak_mem_gb": 2.110217936, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.5922811365849341, "cache_misses": 4606, "cache_hits": 6691, "stall_s": 1.4938332651508972, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3016, "installed": 3016, "late": 29, "dup": 6256, "predictor": "transition2", "m": 8}, "ssd_bytes": 26973831168, "ssd_read_s": 5.189640132477507}38{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:52:28", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 353.465290972963, "prompt_tps": 67.97521311404653, "gen_tps": 39.38169105789814, "decode_tps_wall": 38.76664831243123, "itl_ms_p50": 25.737374904565513, "itl_ms_p95": 36.12379194237292, "itl_ms_p99": 38.14187506213784, "itl_ms_mean": 25.795369048691576, "total_s": 1.9785735410405323, "peak_mem_gb": 1.811898128, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.7109216009303228, "cache_misses": 2983, "cache_hits": 7336, "stall_s": 0.7793248305097222, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1337, "installed": 1337, "late": 0, "dup": 7559, "predictor": "transition2", "m": 8}, "ssd_bytes": 15288238080, "ssd_read_s": 1.9663502274779603}39{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:52:31", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 328.02633300889283, "prompt_tps": 42.735298295847464, "gen_tps": 28.216706367450364, "decode_tps_wall": 27.77589533673794, "itl_ms_p50": 36.09183302614838, "itl_ms_p95": 46.18437506724149, "itl_ms_p99": 49.58812496624887, "itl_ms_mean": 36.0024398089283, "total_s": 2.596180040971376, "peak_mem_gb": 1.740456936, "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": 0.6220268872802482, "cache_misses": 3655, "cache_hits": 6015, "stall_s": 1.1207358334213495, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 4249, "installed": 4246, "late": 22, "dup": 4303, "predictor": "transition2", "m": 8}, "ssd_bytes": 27971813376, "ssd_read_s": 3.889851111569442}40{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:52:33", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 381.94245903287083, "prompt_tps": 47.18371134498646, "gen_tps": 32.91625621517676, "decode_tps_wall": 32.402164007691304, "itl_ms_p50": 30.359124997630715, "itl_ms_p95": 41.354916989803314, "itl_ms_p99": 42.30225004721433, "itl_ms_mean": 30.862136237648507, "total_s": 2.326257042004727, "peak_mem_gb": 1.754967608, "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": 0.9807824203744531, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3271, "installed": 3274, "late": 24, "dup": 5441, "predictor": "transition2", "m": 8}, "ssd_bytes": 23679074304, "ssd_read_s": 3.1471330774948}41{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:52:35", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 387.1852499432862, "prompt_tps": 74.96652363611523, "gen_tps": 34.813347879089356, "decode_tps_wall": 34.26958505075751, "itl_ms_p50": 29.05166696291417, "itl_ms_p95": 37.13020891882479, "itl_ms_p99": 38.455875008367, "itl_ms_mean": 29.180394175151985, "total_s": 2.225550082977861, "peak_mem_gb": 1.912478948, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.6512071762572764, "cache_misses": 3655, "cache_hits": 6824, "stall_s": 0.7946433289907873, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2456, "installed": 2456, "late": 0, "dup": 6464, "predictor": "transition2", "m": 8}, "ssd_bytes": 21626486784, "ssd_read_s": 2.294017414445989}42{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:52:37", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 305.5981670040637, "prompt_tps": 62.255984819796424, "gen_tps": 42.333203310080826, "decode_tps_wall": 41.67207088920581, "itl_ms_p50": 23.547915974631906, "itl_ms_p95": 29.626416973769665, "itl_ms_p99": 34.785165917128325, "itl_ms_mean": 23.9968875715036, "total_s": 1.8174020840087906, "peak_mem_gb": 1.774123092, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.7386317498984978, "cache_misses": 2575, "cache_hits": 7277, "stall_s": 0.5245005063479766, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1242, "installed": 1242, "late": 0, "dup": 7422, "predictor": "transition2", "m": 8}, "ssd_bytes": 13508149248, "ssd_read_s": 1.3246690002270043}43{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:52:39", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 504.75587497930974, "prompt_tps": 63.43570765636268, "gen_tps": 33.63231791437441, "decode_tps_wall": 33.10700722765589, "itl_ms_p50": 29.965166002511978, "itl_ms_p95": 35.564167075790465, "itl_ms_p99": 45.27879203669727, "itl_ms_mean": 30.205085984475556, "total_s": 2.4076762920012698, "peak_mem_gb": 2.090230864, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.5951137470124812, "cache_misses": 4574, "cache_hits": 6723, "stall_s": 0.9355781902559102, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2981, "installed": 2981, "late": 0, "dup": 6403, "predictor": "transition2", "m": 8}, "ssd_bytes": 26736721920, "ssd_read_s": 2.909290548763238}44{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:52:42", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 323.07720894459635, "prompt_tps": 55.77347093293254, "gen_tps": 34.76556916605975, "decode_tps_wall": 34.222504322196336, "itl_ms_p50": 29.257082962431014, "itl_ms_p95": 35.35625000949949, "itl_ms_p99": 37.514208001084626, "itl_ms_mean": 29.220538350590875, "total_s": 2.1639711250318214, "peak_mem_gb": 1.83861416, "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": 0.6556537457320747, "cache_misses": 3429, "cache_hits": 6529, "stall_s": 0.7441158297006041, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3262, "installed": 3262, "late": 1, "dup": 5450, "predictor": "transition2", "m": 8}, "ssd_bytes": 23679074304, "ssd_read_s": 2.4573283984791487}45{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:52:44", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 303.9848329499364, "prompt_tps": 46.10153863994922, "gen_tps": 32.26179979669117, "decode_tps_wall": 31.757873934920628, "itl_ms_p50": 31.53604199178517, "itl_ms_p95": 38.562958012335, "itl_ms_p99": 41.18674993515015, "itl_ms_mean": 31.48825396968436, "total_s": 2.287744833040051, "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": 0.6189245087900724, "cache_misses": 3685, "cache_hits": 5985, "stall_s": 0.812436431995593, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 4272, "installed": 4272, "late": 3, "dup": 4280, "predictor": "transition2", "m": 8}, "ssd_bytes": 28159377408, "ssd_read_s": 2.9446414752164856}46{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:52:46", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 369.5938750170171, "prompt_tps": 78.53176300418164, "gen_tps": 35.86073790447289, "decode_tps_wall": 35.30060591962893, "itl_ms_p50": 28.571082977578044, "itl_ms_p95": 35.74508405290544, "itl_ms_p99": 37.75250003673136, "itl_ms_mean": 28.328125649649238, "total_s": 2.154265790944919, "peak_mem_gb": 1.925994748, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.6514934631167096, "cache_misses": 3652, "cache_hits": 6827, "stall_s": 0.7625655947485939, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2452, "installed": 2452, "late": 0, "dup": 6468, "predictor": "transition2", "m": 8}, "ssd_bytes": 21601714176, "ssd_read_s": 2.2692865426652133}47{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:52:48", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 354.6893330058083, "prompt_tps": 67.72682667127343, "gen_tps": 41.44801108449971, "decode_tps_wall": 40.80068428436811, "itl_ms_p50": 23.674833006225526, "itl_ms_p95": 30.851582996547222, "itl_ms_p99": 33.191584050655365, "itl_ms_mean": 24.509392857980277, "total_s": 1.8987810830585659, "peak_mem_gb": 1.936432592, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.7095648803178602, "cache_misses": 2997, "cache_hits": 7322, "stall_s": 0.5869833705946803, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1354, "installed": 1354, "late": 0, "dup": 7542, "predictor": "transition2", "m": 8}, "ssd_bytes": 15397945344, "ssd_read_s": 1.528933696099557}48{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:52:50", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 293.2850420475006, "prompt_tps": 64.85182984578265, "gen_tps": 42.035485166208865, "decode_tps_wall": 41.378936643178235, "itl_ms_p50": 22.92745898012072, "itl_ms_p95": 29.90829199552536, "itl_ms_p99": 33.40404201298952, "itl_ms_mean": 24.166884920781566, "total_s": 1.8157987920567393, "peak_mem_gb": 1.77871062, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.7416768168899716, "cache_misses": 2545, "cache_hits": 7307, "stall_s": 0.5172337950207293, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1220, "installed": 1220, "late": 0, "dup": 7444, "predictor": "transition2", "m": 8}, "ssd_bytes": 13324124160, "ssd_read_s": 1.2893915047170594}49{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:52:52", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 378.07437498122454, "prompt_tps": 76.7743733148134, "gen_tps": 35.166428437289746, "decode_tps_wall": 34.61712498277362, "itl_ms_p50": 29.142415965907276, "itl_ms_p95": 36.59095801413059, "itl_ms_p99": 38.46429195255041, "itl_ms_mean": 28.887436507151477, "total_s": 2.1979828749317676, "peak_mem_gb": 1.913297764, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.6513980341635652, "cache_misses": 3653, "cache_hits": 6826, "stall_s": 0.7939919838681817, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2486, "installed": 2486, "late": 0, "dup": 6434, "predictor": "transition2", "m": 8}, "ssd_bytes": 21725577216, "ssd_read_s": 2.304772370494902}50{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:52:54", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 302.420417079702, "prompt_tps": 62.89671670966046, "gen_tps": 42.07464256237184, "decode_tps_wall": 41.41748720225738, "itl_ms_p50": 23.17712502554059, "itl_ms_p95": 29.35124991927296, "itl_ms_p99": 35.00116697978228, "itl_ms_mean": 24.144390873270968, "total_s": 1.823517042095773, "peak_mem_gb": 1.774123092, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.7386317498984978, "cache_misses": 2575, "cache_hits": 7277, "stall_s": 0.5174110381631181, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1242, "installed": 1242, "late": 0, "dup": 7422, "predictor": "transition2", "m": 8}, "ssd_bytes": 13508149248, "ssd_read_s": 1.2985591932665557}51{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:52:56", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 323.702875059098, "prompt_tps": 55.65837616338047, "gen_tps": 33.72026162156194, "decode_tps_wall": 33.19346415416032, "itl_ms_p50": 29.05595803167671, "itl_ms_p95": 37.15220803860575, "itl_ms_p99": 38.89787499792874, "itl_ms_mean": 30.12641269846686, "total_s": 2.2216668750625104, "peak_mem_gb": 1.832781476, "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": 0.6555533239606347, "cache_misses": 3430, "cache_hits": 6528, "stall_s": 0.7704306307714432, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3225, "installed": 3224, "late": 0, "dup": 5487, "predictor": "transition2", "m": 8}, "ssd_bytes": 23551672320, "ssd_read_s": 2.4809985145693645}52{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:52:58", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 381.26549997832626, "prompt_tps": 62.9984231902653, "gen_tps": 41.532063736529814, "decode_tps_wall": 40.88342042071837, "itl_ms_p50": 23.87808298226446, "itl_ms_p95": 30.905375024303794, "itl_ms_p99": 33.17204106133431, "itl_ms_mean": 24.4597929847678, "total_s": 1.9222324580186978, "peak_mem_gb": 1.94711496, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.7108246923151468, "cache_misses": 2984, "cache_hits": 7335, "stall_s": 0.6018826133804396, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1374, "installed": 1375, "late": 0, "dup": 7522, "predictor": "transition2", "m": 8}, "ssd_bytes": 15422717952, "ssd_read_s": 1.5850066980347037}53{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:53:00", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 296.582916053012, "prompt_tps": 47.25632718283334, "gen_tps": 31.66000322566727, "decode_tps_wall": 31.165460211841552, "itl_ms_p50": 31.57820808701217, "itl_ms_p95": 40.186041966080666, "itl_ms_p99": 41.83170909527689, "itl_ms_mean": 32.08680357044888, "total_s": 2.3180515409912914, "peak_mem_gb": 1.765490084, "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": 0.6216132368148914, "cache_misses": 3659, "cache_hits": 6011, "stall_s": 0.8197586905444041, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 4262, "installed": 4262, "late": 2, "dup": 4290, "predictor": "transition2", "m": 8}, "ssd_bytes": 28031975424, "ssd_read_s": 2.9535520400386304}54{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:53:03", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 469.61216698400676, "prompt_tps": 68.18820478879253, "gen_tps": 34.05711275171166, "decode_tps_wall": 33.52519782793929, "itl_ms_p50": 29.667209018953145, "itl_ms_p95": 36.3955000648275, "itl_ms_p99": 45.34125002101064, "itl_ms_mean": 29.828310190212157, "total_s": 2.3487957089673728, "peak_mem_gb": 2.1066804, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.5950252279366204, "cache_misses": 4575, "cache_hits": 6722, "stall_s": 0.9184405328705907, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2953, "installed": 2953, "late": 0, "dup": 6431, "predictor": "transition2", "m": 8}, "ssd_bytes": 26641170432, "ssd_read_s": 2.8397510712966323}55{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:53:11", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 753.2049170695245, "prompt_tps": 42.64084524581707, "gen_tps": 27.340811820146158, "decode_tps_wall": 26.91468525848031, "itl_ms_p50": 36.52970807161182, "itl_ms_p95": 46.35674995370209, "itl_ms_p99": 52.91012499947101, "itl_ms_mean": 37.1544378244185, "total_s": 3.09393450000789, "peak_mem_gb": 2.110217936, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.5786491989023634, "cache_misses": 4760, "cache_hits": 6537, "stall_s": 1.5305998735129833, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2771, "installed": 2771, "late": 36, "dup": 6381, "predictor": "transition3", "m": 8}, "ssd_bytes": 26651787264, "ssd_read_s": 5.400361988926306}56{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:53:13", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 420.8150419872254, "prompt_tps": 57.087344311843424, "gen_tps": 34.54763082527328, "decode_tps_wall": 34.00806503975473, "itl_ms_p50": 28.49091694224626, "itl_ms_p95": 38.611458032391965, "itl_ms_p99": 40.60274991206825, "itl_ms_mean": 29.40478968241858, "total_s": 2.273316791979596, "peak_mem_gb": 1.943576016, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.7041379978680105, "cache_misses": 3053, "cache_hits": 7266, "stall_s": 0.9600879641948268, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1204, "installed": 1203, "late": 6, "dup": 7692, "predictor": "transition3", "m": 8}, "ssd_bytes": 15065284608, "ssd_read_s": 2.7973526313435286}57{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:53:16", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 365.8355000661686, "prompt_tps": 38.30144935160632, "gen_tps": 25.917288957690737, "decode_tps_wall": 25.512444964646704, "itl_ms_p50": 39.969999925233424, "itl_ms_p95": 46.09837499447167, "itl_ms_p99": 51.19445803575218, "itl_ms_mean": 39.19655687197866, "total_s": 2.835218583000824, "peak_mem_gb": 1.75032922, "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": 0.6042399172699069, "cache_misses": 3827, "cache_hits": 5843, "stall_s": 1.285978739731945, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3844, "installed": 3843, "late": 91, "dup": 4708, "predictor": "transition3", "m": 8}, "ssd_bytes": 27147239424, "ssd_read_s": 5.100391939282417}58{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:53:18", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 383.7739999871701, "prompt_tps": 46.95035189643236, "gen_tps": 28.602661656534913, "decode_tps_wall": 28.15585883307409, "itl_ms_p50": 36.13400005269796, "itl_ms_p95": 43.767041992396116, "itl_ms_p99": 47.90783300995827, "itl_ms_mean": 35.51658665177427, "total_s": 2.621318959048949, "peak_mem_gb": 1.85630888, "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": 0.6438039767021491, "cache_misses": 3547, "cache_hits": 6411, "stall_s": 1.1678438034141436, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3008, "installed": 3006, "late": 108, "dup": 5704, "predictor": "transition3", "m": 8}, "ssd_bytes": 23197777920, "ssd_read_s": 4.267102239537053}59{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:53:21", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 444.09462506882846, "prompt_tps": 65.35459799267639, "gen_tps": 29.238124163006884, "decode_tps_wall": 28.781462001861236, "itl_ms_p50": 35.20095895510167, "itl_ms_p95": 44.54370797611773, "itl_ms_p99": 45.570999966003, "itl_ms_mean": 34.74458663480445, "total_s": 2.633003583061509, "peak_mem_gb": 1.919294376, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.6341253936444318, "cache_misses": 3834, "cache_hits": 6645, "stall_s": 1.2079252713592723, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2304, "installed": 2308, "late": 34, "dup": 6616, "predictor": "transition3", "m": 8}, "ssd_bytes": 21722038272, "ssd_read_s": 4.141525521525182}60{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:53:23", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 349.29945797193795, "prompt_tps": 54.44807314196366, "gen_tps": 35.810574934728145, "decode_tps_wall": 35.25124427394701, "itl_ms_p50": 26.920875068753958, "itl_ms_p95": 36.960417055524886, "itl_ms_p99": 37.821333040483296, "itl_ms_mean": 28.3677929842342, "total_s": 2.1364704159786925, "peak_mem_gb": 1.785985096, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.7290905399918798, "cache_misses": 2669, "cache_hits": 7183, "stall_s": 0.860856703016907, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1155, "installed": 1155, "late": 9, "dup": 7509, "predictor": "transition3", "m": 8}, "ssd_bytes": 13532921856, "ssd_read_s": 2.4766576415859163}61{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:53:26", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 613.1644169799984, "prompt_tps": 52.21898744345239, "gen_tps": 27.392022640473126, "decode_tps_wall": 26.96422280617183, "itl_ms_p50": 36.512749968096614, "itl_ms_p95": 44.91145897191018, "itl_ms_p99": 56.61400000099093, "itl_ms_mean": 37.086179237886675, "total_s": 2.9495937089668587, "peak_mem_gb": 2.097308752, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.5817473665574932, "cache_misses": 4725, "cache_hits": 6572, "stall_s": 1.4236707100644708, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2720, "installed": 2720, "late": 39, "dup": 6664, "predictor": "transition3", "m": 8}, "ssd_bytes": 26347438080, "ssd_read_s": 4.9407663334859535}62{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:53:29", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 371.72820803243667, "prompt_tps": 48.46693041974209, "gen_tps": 28.816363362515276, "decode_tps_wall": 28.36622317108597, "itl_ms_p50": 35.94166599214077, "itl_ms_p95": 43.90633397269994, "itl_ms_p99": 47.56720794830471, "itl_ms_mean": 35.25319511056065, "total_s": 2.5926794999977574, "peak_mem_gb": 1.842153112, "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": 0.6442056637879092, "cache_misses": 3543, "cache_hits": 6415, "stall_s": 1.1450044836383313, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2992, "installed": 2989, "late": 97, "dup": 5720, "predictor": "transition3", "m": 8}, "ssd_bytes": 23126999040, "ssd_read_s": 4.172673489316367}63{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:53:32", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 354.6910840086639, "prompt_tps": 39.50963831692955, "gen_tps": 25.957192597923797, "decode_tps_wall": 25.551733243549446, "itl_ms_p50": 39.379875059239566, "itl_ms_p95": 46.66274995543063, "itl_ms_p99": 49.57350005861372, "itl_ms_mean": 39.13628834758013, "total_s": 2.820277249906212, "peak_mem_gb": 1.766499812, "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": 0.6010341261633919, "cache_misses": 3858, "cache_hits": 5812, "stall_s": 1.2766682531218976, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3841, "installed": 3839, "late": 84, "dup": 4711, "predictor": "transition3", "m": 8}, "ssd_bytes": 27246329856, "ssd_read_s": 5.046303864219226}64{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:53:34", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 467.11124991998076, "prompt_tps": 62.131434986674485, "gen_tps": 29.35151327800442, "decode_tps_wall": 28.893030604599495, "itl_ms_p50": 35.02037504222244, "itl_ms_p95": 42.79870796017349, "itl_ms_p99": 44.56633294466883, "itl_ms_mean": 34.61042262007675, "total_s": 2.6475678749848157, "peak_mem_gb": 1.923137056, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.6344116805038649, "cache_misses": 3831, "cache_hits": 6648, "stall_s": 1.2043944289907813, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2302, "installed": 2307, "late": 33, "dup": 6618, "predictor": "transition3", "m": 8}, "ssd_bytes": 21704343552, "ssd_read_s": 4.134974777814932}65{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:53:36", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 410.85666709113866, "prompt_tps": 58.46602798684686, "gen_tps": 34.395164607205125, "decode_tps_wall": 33.857986572159255, "itl_ms_p50": 28.66954100318253, "itl_ms_p95": 38.134667091071606, "itl_ms_p99": 40.19066703040153, "itl_ms_mean": 29.5351289678365, "total_s": 2.271569792064838, "peak_mem_gb": 1.939971536, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.7022967341796685, "cache_misses": 3072, "cache_hits": 7247, "stall_s": 0.96135191002395, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1209, "installed": 1208, "late": 8, "dup": 7687, "predictor": "transition3", "m": 8}, "ssd_bytes": 15150219264, "ssd_read_s": 2.8135463966755196}66{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:53:39", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 323.7640829756856, "prompt_tps": 58.73757487780616, "gen_tps": 37.34104801238553, "decode_tps_wall": 36.757786275582944, "itl_ms_p50": 26.66179195512086, "itl_ms_p95": 35.13124992605299, "itl_ms_p99": 36.5643750410527, "itl_ms_mean": 27.20512036559364, "total_s": 2.037686666008085, "peak_mem_gb": 1.784448924, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.7333536337799431, "cache_misses": 2627, "cache_hits": 7225, "stall_s": 0.8013249097857624, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1137, "installed": 1138, "late": 7, "dup": 7527, "predictor": "transition3", "m": 8}, "ssd_bytes": 13320585216, "ssd_read_s": 2.287489281501621}67{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:53:41", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 440.6553329899907, "prompt_tps": 65.85454881905765, "gen_tps": 29.704000209209045, "decode_tps_wall": 29.24005841482512, "itl_ms_p50": 34.30370893329382, "itl_ms_p95": 42.71358298137784, "itl_ms_p99": 44.67908293008804, "itl_ms_mean": 34.1996580791024, "total_s": 2.595233791973442, "peak_mem_gb": 1.92016208, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.6337436778318541, "cache_misses": 3838, "cache_hits": 6641, "stall_s": 1.181948950397782, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2326, "installed": 2326, "late": 33, "dup": 6594, "predictor": "transition3", "m": 8}, "ssd_bytes": 21814050816, "ssd_read_s": 4.06036287418101}68{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:53:43", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 352.86379198078066, "prompt_tps": 53.89972802061477, "gen_tps": 36.22907587479348, "decode_tps_wall": 35.663209385743905, "itl_ms_p50": 26.90387493930757, "itl_ms_p95": 35.17370799090713, "itl_ms_p99": 39.51866598799825, "itl_ms_mean": 28.04010119178288, "total_s": 2.119390167063102, "peak_mem_gb": 1.787987868, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.729192042224929, "cache_misses": 2668, "cache_hits": 7184, "stall_s": 0.8518375136191025, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1154, "installed": 1154, "late": 10, "dup": 7510, "predictor": "transition3", "m": 8}, "ssd_bytes": 13525843968, "ssd_read_s": 2.446215729578398}69{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:53:46", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 378.37908300571144, "prompt_tps": 47.60749737789905, "gen_tps": 28.540004306655902, "decode_tps_wall": 28.09421289530375, "itl_ms_p50": 35.74779094196856, "itl_ms_p95": 45.13195808976889, "itl_ms_p99": 48.81591710727662, "itl_ms_mean": 35.594519188938044, "total_s": 2.620833791908808, "peak_mem_gb": 1.832715924, "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": 0.643904398473589, "cache_misses": 3546, "cache_hits": 6412, "stall_s": 1.1633029583608732, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2958, "installed": 2954, "late": 106, "dup": 5754, "predictor": "transition3", "m": 8}, "ssd_bytes": 23017291776, "ssd_read_s": 4.207141849328764}70{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:53:48", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 406.8406249862164, "prompt_tps": 59.0398097534521, "gen_tps": 35.0150041351082, "decode_tps_wall": 34.46810605404178, "itl_ms_p50": 28.314916999079287, "itl_ms_p95": 38.30704092979431, "itl_ms_p99": 38.786791963502765, "itl_ms_mean": 29.012328047039254, "total_s": 2.2346172919496894, "peak_mem_gb": 1.916378552, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.7030720031010756, "cache_misses": 3064, "cache_hits": 7255, "stall_s": 0.9412726262817159, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1228, "installed": 1232, "late": 8, "dup": 7668, "predictor": "transition3", "m": 8}, "ssd_bytes": 15189147648, "ssd_read_s": 2.764597364817746}71{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:53:51", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 347.69745799712837, "prompt_tps": 40.29696456872564, "gen_tps": 26.761941039115683, "decode_tps_wall": 26.343887607006273, "itl_ms_p50": 38.27670798636973, "itl_ms_p95": 45.40520906448364, "itl_ms_p99": 50.906124990433455, "itl_ms_mean": 37.95946957100006, "total_s": 2.7391440409701318, "peak_mem_gb": 1.748074372, "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": 0.6040330920372285, "cache_misses": 3829, "cache_hits": 5841, "stall_s": 1.2457460204605013, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3847, "installed": 3842, "late": 90, "dup": 4705, "predictor": "transition3", "m": 8}, "ssd_bytes": 27164934144, "ssd_read_s": 4.955624026013538}72{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:53:54", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 589.582083048299, "prompt_tps": 54.30586079041611, "gen_tps": 27.95022901670601, "decode_tps_wall": 27.5137635278525, "itl_ms_p50": 36.157459020614624, "itl_ms_p95": 44.21754099894315, "itl_ms_p99": 51.41895799897611, "itl_ms_mean": 36.34544576163448, "total_s": 2.879345166031271, "peak_mem_gb": 2.08730024, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.5817473665574932, "cache_misses": 4725, "cache_hits": 6572, "stall_s": 1.3919748768676072, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2695, "installed": 2700, "late": 37, "dup": 6689, "predictor": "transition3", "m": 8}, "ssd_bytes": 26258964480, "ssd_read_s": 4.873588268761523}73{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:54:01", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 738.4890830144286, "prompt_tps": 43.654311526418894, "gen_tps": 19.634206300194876, "decode_tps_wall": 19.32782107607674, "itl_ms_p50": 51.67612503282726, "itl_ms_p95": 67.43425002787262, "itl_ms_p99": 68.57829191721976, "itl_ms_mean": 51.738889555313754, "total_s": 3.9980391249991953, "peak_mem_gb": 2.090229456, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.5930778082676816, "cache_misses": 4597, "cache_hits": 6700, "stall_s": 1.8813947818707675, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 11228, "installed": 10981, "late": 500, "dup": 5412, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 55377395712, "ssd_read_s": 8.49053135048598}74{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:54:05", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 488.09512495063245, "prompt_tps": 49.23585268689506, "gen_tps": 21.883011480968356, "decode_tps_wall": 21.54128737238935, "itl_ms_p50": 46.18499998468906, "itl_ms_p95": 56.120374938473105, "itl_ms_p99": 59.813000028952956, "itl_ms_mean": 46.42248082543827, "total_s": 3.4127114169532433, "peak_mem_gb": 2.604126676, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.7457117937784669, "cache_misses": 2624, "cache_hits": 7695, "stall_s": 1.4655210030032322, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 10649, "installed": 10695, "late": 639, "dup": 5991, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 46926397440, "ssd_read_s": 6.642568752751686}75{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:54:08", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 396.3068750454113, "prompt_tps": 35.381113498651956, "gen_tps": 19.234880529914122, "decode_tps_wall": 18.93443605876646, "itl_ms_p50": 53.189291968010366, "itl_ms_p95": 62.700584065169096, "itl_ms_p99": 68.87941702734679, "itl_ms_mean": 52.813825397086994, "total_s": 3.723577875061892, "peak_mem_gb": 2.188492224, "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": 0.609100310237849, "cache_misses": 3780, "cache_hits": 5890, "stall_s": 1.6934111309237778, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 11309, "installed": 11200, "late": 522, "dup": 5331, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 53455749120, "ssd_read_s": 8.21669495117385}76{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:54:12", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 491.03087501134723, "prompt_tps": 36.704659129798834, "gen_tps": 20.09872583298128, "decode_tps_wall": 19.784749257624057, "itl_ms_p50": 51.59587494563311, "itl_ms_p95": 60.76370901428163, "itl_ms_p99": 66.465082927607, "itl_ms_mean": 50.54398147677559, "total_s": 3.6753017080482095, "peak_mem_gb": 2.802191004, "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": 0.6431010243020687, "cache_misses": 3554, "cache_hits": 6404, "stall_s": 1.6476899560075253, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 10812, "installed": 10927, "late": 570, "dup": 5828, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 50773229568, "ssd_read_s": 7.653688264661469}77{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:54:16", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 472.8152919560671, "prompt_tps": 61.41383647956028, "gen_tps": 18.834992443979438, "decode_tps_wall": 18.540842788427884, "itl_ms_p50": 54.68833306804299, "itl_ms_p95": 64.29741601459682, "itl_ms_p99": 75.55237493943423, "itl_ms_mean": 53.9349808102651, "total_s": 3.8707190830027685, "peak_mem_gb": 2.420217636, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.6525431816012979, "cache_misses": 3641, "cache_hits": 6838, "stall_s": 1.6587911994429305, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 11375, "installed": 11344, "late": 481, "dup": 5265, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 53162016768, "ssd_read_s": 7.974729008390568}78{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:54:19", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 415.41041689924896, "prompt_tps": 45.806256416292655, "gen_tps": 20.88992012176414, "decode_tps_wall": 20.563628660998237, "itl_ms_p50": 49.125874997116625, "itl_ms_p95": 58.533707982860506, "itl_ms_p99": 59.69133402686566, "itl_ms_mean": 48.62954960359881, "total_s": 3.479072041925974, "peak_mem_gb": 2.352155628, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.7481729598051157, "cache_misses": 2481, "cache_hits": 7371, "stall_s": 1.4357508576940745, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 10667, "installed": 10665, "late": 554, "dup": 5973, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 46438023168, "ssd_read_s": 6.680345768225379}79{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:54:24", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 649.8539160238579, "prompt_tps": 49.28818678012778, "gen_tps": 18.667066904812263, "decode_tps_wall": 18.37546700941863, "itl_ms_p50": 54.8915839754045, "itl_ms_p95": 64.01470908895135, "itl_ms_p99": 74.58949997089803, "itl_ms_mean": 54.4203855873396, "total_s": 4.078338208026253, "peak_mem_gb": 2.717506984, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.6018411967779056, "cache_misses": 4498, "cache_hits": 6799, "stall_s": 1.8833149742567912, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 11217, "installed": 11188, "late": 535, "dup": 5423, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 55741906944, "ssd_read_s": 8.682453730027191}80{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:54:27", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 485.24133302271366, "prompt_tps": 37.14527554034433, "gen_tps": 19.30410680855894, "decode_tps_wall": 19.002615310633526, "itl_ms_p50": 54.14712498895824, "itl_ms_p95": 60.98545901477337, "itl_ms_p99": 64.66695899143815, "itl_ms_mean": 52.624335316645485, "total_s": 3.800574457971379, "peak_mem_gb": 2.607663772, "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": 0.6432014460735087, "cache_misses": 3553, "cache_hits": 6405, "stall_s": 1.6771054364508018, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 10876, "installed": 10940, "late": 549, "dup": 5764, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 50989105152, "ssd_read_s": 7.8780766791896895}81{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:54:31", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 406.9164589745924, "prompt_tps": 34.45566857049648, "gen_tps": 18.40166272644055, "decode_tps_wall": 18.114272594287947, "itl_ms_p50": 55.84937508683652, "itl_ms_p95": 64.01233398355544, "itl_ms_p99": 73.2377499807626, "itl_ms_mean": 55.205087303110055, "total_s": 3.884836959070526, "peak_mem_gb": 2.196848, "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": 0.6103412616339193, "cache_misses": 3768, "cache_hits": 5902, "stall_s": 1.6982223885133862, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 11411, "installed": 11382, "late": 513, "dup": 5229, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 53664546816, "ssd_read_s": 8.464784750132822}82{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:54:35", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 493.30812506377697, "prompt_tps": 58.85286129871006, "gen_tps": 18.823585699780406, "decode_tps_wall": 18.529541651272485, "itl_ms_p50": 53.91395802143961, "itl_ms_p95": 62.34887498430908, "itl_ms_p99": 69.93033294565976, "itl_ms_mean": 53.96787566687202, "total_s": 3.8932842920767143, "peak_mem_gb": 2.48598714, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.6518751789292871, "cache_misses": 3648, "cache_hits": 6831, "stall_s": 1.72785309352912, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 11293, "installed": 11307, "late": 529, "dup": 5347, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 52882440192, "ssd_read_s": 8.228081482113339}83{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:54:39", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 467.34316705260426, "prompt_tps": 51.4308995639455, "gen_tps": 21.603713406163724, "decode_tps_wall": 21.266237944446882, "itl_ms_p50": 46.558125060983, "itl_ms_p95": 56.56345805618912, "itl_ms_p99": 59.992791968397796, "itl_ms_mean": 47.02289152469131, "total_s": 3.429785333108157, "peak_mem_gb": 2.412978448, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.7453241593177633, "cache_misses": 2628, "cache_hits": 7691, "stall_s": 1.4848567594308406, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 10669, "installed": 10670, "late": 620, "dup": 5971, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 47067955200, "ssd_read_s": 6.713389698998071}84{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:54:42", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 406.449917005375, "prompt_tps": 46.81849823455439, "gen_tps": 19.356718769709314, "decode_tps_wall": 19.054455895962505, "itl_ms_p50": 52.89687495678663, "itl_ms_p95": 64.27420896943659, "itl_ms_p99": 66.87500001862645, "itl_ms_mean": 52.4811626981116, "total_s": 3.712763166986406, "peak_mem_gb": 2.26417756, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.7502030044660982, "cache_misses": 2461, "cache_hits": 7391, "stall_s": 1.4805924523388967, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 10686, "installed": 10673, "late": 527, "dup": 5954, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 46484029440, "ssd_read_s": 7.116914416081272}85{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:54:46", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 526.9134580157697, "prompt_tps": 55.12156562908391, "gen_tps": 18.86589467621387, "decode_tps_wall": 18.571189200660164, "itl_ms_p50": 54.11991698201746, "itl_ms_p95": 61.992709059268236, "itl_ms_p99": 73.37999995797873, "itl_ms_mean": 53.846847888688366, "total_s": 3.9192648750031367, "peak_mem_gb": 2.463901516, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.6499665998663995, "cache_misses": 3668, "cache_hits": 6811, "stall_s": 1.7454534453572705, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 11277, "installed": 11278, "late": 534, "dup": 5363, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 52917829632, "ssd_read_s": 8.160517507814802}86{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:54:50", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 396.0263750050217, "prompt_tps": 48.05188077740327, "gen_tps": 21.062078045370882, "decode_tps_wall": 20.73320511655182, "itl_ms_p50": 48.925124923698604, "itl_ms_p95": 58.50166699383408, "itl_ms_p99": 62.86554201506078, "itl_ms_mean": 48.2318095238288, "total_s": 3.434630375006236, "peak_mem_gb": 2.314738664, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.747563946406821, "cache_misses": 2487, "cache_hits": 7365, "stall_s": 1.4233429500600323, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 10604, "installed": 10594, "late": 553, "dup": 6036, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 46278770688, "ssd_read_s": 6.7789444053778425}87{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:54:53", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 445.88733301497996, "prompt_tps": 40.42741134778498, "gen_tps": 18.957355326548118, "decode_tps_wall": 18.661272635469416, "itl_ms_p50": 55.01899996306747, "itl_ms_p95": 64.2057079821825, "itl_ms_p99": 71.09279197175056, "itl_ms_mean": 53.58691336513157, "total_s": 3.821862875018269, "peak_mem_gb": 2.357238344, "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": 0.6404900582446275, "cache_misses": 3580, "cache_hits": 6378, "stall_s": 1.6816431327024475, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 10842, "installed": 10872, "late": 565, "dup": 5798, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 51074039808, "ssd_read_s": 7.943893982679583}88{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:54:57", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 467.50991605222225, "prompt_tps": 51.400003202950664, "gen_tps": 20.63356794043737, "decode_tps_wall": 20.311353708294387, "itl_ms_p50": 50.07370898965746, "itl_ms_p95": 59.30404202081263, "itl_ms_p99": 69.94429090991616, "itl_ms_mean": 49.23354761881961, "total_s": 3.5692234160378575, "peak_mem_gb": 2.388152764, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.7461963368543464, "cache_misses": 2619, "cache_hits": 7700, "stall_s": 1.5268688319483772, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 10645, "installed": 10638, "late": 624, "dup": 5995, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 46929936384, "ssd_read_s": 6.952907785656862}89{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:55:01", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 401.94941696245223, "prompt_tps": 34.885749172632515, "gen_tps": 18.349128659023847, "decode_tps_wall": 18.062554279328474, "itl_ms_p50": 55.842416011728346, "itl_ms_p95": 64.86345804296434, "itl_ms_p99": 80.04754199646413, "itl_ms_mean": 55.3631554283793, "total_s": 3.889828208950348, "peak_mem_gb": 2.192671264, "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": 0.6107549120992761, "cache_misses": 3764, "cache_hits": 5906, "stall_s": 1.7141088997013867, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 11404, "installed": 11396, "late": 496, "dup": 5236, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 53625618432, "ssd_read_s": 8.446497321594507}90{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:55:05", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 672.1777920611203, "prompt_tps": 47.64901139788459, "gen_tps": 19.377153200391433, "decode_tps_wall": 19.07446627689362, "itl_ms_p50": 53.06133395060897, "itl_ms_p95": 61.16104102693498, "itl_ms_p99": 80.60945908073336, "itl_ms_mean": 52.42610647572234, "total_s": 3.9750225000316277, "peak_mem_gb": 2.717506952, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.6066212268743915, "cache_misses": 4444, "cache_hits": 6853, "stall_s": 1.8803486177930608, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 11176, "installed": 11143, "late": 555, "dup": 5464, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 55363239936, "ssd_read_s": 8.55122381716501}91{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:55:13", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 776.9622500054538, "prompt_tps": 41.34613986780918, "gen_tps": 25.967484486226184, "decode_tps_wall": 25.56300531981259, "itl_ms_p50": 38.845291011966765, "itl_ms_p95": 48.301166971214116, "itl_ms_p99": 68.34666698705405, "itl_ms_mean": 39.119031095492936, "total_s": 3.2414612090215087, "peak_mem_gb": 2.148032208, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.6855802425422679, "cache_misses": 3552, "cache_hits": 7745, "stall_s": 1.4738099195528775, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 6898, "installed": 6866, "late": 655, "dup": 11126, "predictor": "transition2", "m": 16}, "ssd_bytes": 36981964800, "ssd_read_s": 6.262643188936636}92{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:55:16", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 414.5083329640329, "prompt_tps": 57.9639732617623, "gen_tps": 31.666907672355048, "decode_tps_wall": 31.172245270645256, "itl_ms_p50": 31.223458936437964, "itl_ms_p95": 41.79754201322794, "itl_ms_p99": 47.1361669478938, "itl_ms_mean": 32.079819445720034, "total_s": 2.435536958044395, "peak_mem_gb": 1.964744136, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.7671285977323384, "cache_misses": 2403, "cache_hits": 7916, "stall_s": 0.8962502006907016, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 4647, "installed": 4677, "late": 282, "dup": 12545, "predictor": "transition2", "m": 16}, "ssd_bytes": 24949555200, "ssd_read_s": 4.040312738739885}93{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:55:19", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 369.6352079277858, "prompt_tps": 37.916168027005675, "gen_tps": 22.871121466101666, "decode_tps_wall": 22.51386579064535, "itl_ms_p50": 44.827249948866665, "itl_ms_p95": 54.21820795163512, "itl_ms_p99": 68.7379160663113, "itl_ms_mean": 44.41707209676563, "total_s": 3.1679107500240207, "peak_mem_gb": 1.781939684, "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": 0.7026887280248191, "cache_misses": 2875, "cache_hits": 6795, "stall_s": 1.3573551709996536, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 8061, "installed": 7991, "late": 741, "dup": 8475, "predictor": "transition2", "m": 16}, "ssd_bytes": 38701891584, "ssd_read_s": 6.064778626896441}94{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:55:22", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 423.7950840033591, "prompt_tps": 42.51287290038875, "gen_tps": 26.07871480406025, "decode_tps_wall": 25.671433636822613, "itl_ms_p50": 39.0990839805454, "itl_ms_p95": 49.68541697598994, "itl_ms_p99": 55.28849991969764, "itl_ms_mean": 38.953804222512105, "total_s": 2.8778847500216216, "peak_mem_gb": 2.011277752, "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": 0.747941353685479, "cache_misses": 2510, "cache_hits": 7448, "stall_s": 1.1847929114010185, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 6685, "installed": 6714, "late": 727, "dup": 10155, "predictor": "transition2", "m": 16}, "ssd_bytes": 32540590080, "ssd_read_s": 5.085203237365931}95{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:55:25", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 467.8872500080615, "prompt_tps": 62.04282407348515, "gen_tps": 26.945721894644805, "decode_tps_wall": 26.524822949824582, "itl_ms_p50": 38.14550000242889, "itl_ms_p95": 47.59520792867988, "itl_ms_p99": 51.14270804915577, "itl_ms_mean": 37.70053439721879, "total_s": 2.8430209170328453, "peak_mem_gb": 2.01832226, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.7263097623819067, "cache_misses": 2868, "cache_hits": 7611, "stall_s": 1.177170201903209, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 6216, "installed": 6228, "late": 504, "dup": 11024, "predictor": "transition2", "m": 16}, "ssd_bytes": 32147767296, "ssd_read_s": 5.041412454214878}96{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:55:27", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 368.53945802431554, "prompt_tps": 51.6192014373395, "gen_tps": 32.01257493891522, "decode_tps_wall": 31.51259846961477, "itl_ms_p50": 30.93004203401506, "itl_ms_p95": 42.08649997599423, "itl_ms_p99": 46.378000057302415, "itl_ms_mean": 31.733339951772773, "total_s": 2.3677398749860004, "peak_mem_gb": 1.894558416, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.7973000406008932, "cache_misses": 1997, "cache_hits": 7855, "stall_s": 0.8341774244327098, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 4644, "installed": 4642, "late": 369, "dup": 12100, "predictor": "transition2", "m": 16}, "ssd_bytes": 23502127104, "ssd_read_s": 3.8159793339436874}97{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:55:30", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 636.0643749358132, "prompt_tps": 50.34029512861154, "gen_tps": 24.651099877657906, "decode_tps_wall": 24.266084157941314, "itl_ms_p50": 41.11083399038762, "itl_ms_p95": 51.034624921157956, "itl_ms_p99": 73.67904204875231, "itl_ms_mean": 41.20978042815945, "total_s": 3.2322805419098586, "peak_mem_gb": 2.200612016, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.6885013720456759, "cache_misses": 3519, "cache_hits": 7778, "stall_s": 1.4014906310476363, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 6816, "installed": 6815, "late": 646, "dup": 11320, "predictor": "transition2", "m": 16}, "ssd_bytes": 36574986240, "ssd_read_s": 5.90382578340359}98{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:55:33", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 420.7969999406487, "prompt_tps": 42.82554626658236, "gen_tps": 25.505216667393704, "decode_tps_wall": 25.106821885091115, "itl_ms_p50": 38.832667050883174, "itl_ms_p95": 51.45812500268221, "itl_ms_p99": 61.62699998822063, "itl_ms_mean": 39.829812175224696, "total_s": 2.9300751669798046, "peak_mem_gb": 1.87839452, "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": 0.7467362924281984, "cache_misses": 2522, "cache_hits": 7436, "stall_s": 1.146633548894897, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 6759, "installed": 6746, "late": 663, "dup": 10081, "predictor": "transition2", "m": 16}, "ssd_bytes": 32844939264, "ssd_read_s": 5.040094618801959}99{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:55:37", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 400.48387506976724, "prompt_tps": 34.99180972369275, "gen_tps": 21.264402708984, "decode_tps_wall": 20.932234221907454, "itl_ms_p50": 48.412750009447336, "itl_ms_p95": 56.89112504478544, "itl_ms_p99": 75.21420798730105, "itl_ms_mean": 47.77320898470602, "total_s": 3.4101960411062464, "peak_mem_gb": 1.874141156, "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": 0.703722854188211, "cache_misses": 2865, "cache_hits": 6805, "stall_s": 1.42131144576706, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 8055, "installed": 8027, "late": 726, "dup": 8481, "predictor": "transition2", "m": 16}, "ssd_bytes": 38645268480, "ssd_read_s": 6.298445483553223}100{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:55:40", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 488.34279098082334, "prompt_tps": 59.44349613376352, "gen_tps": 25.35182381699112, "decode_tps_wall": 24.955831723044728, "itl_ms_p50": 39.084124960936606, "itl_ms_p95": 51.17345799226314, "itl_ms_p99": 53.56079200282693, "itl_ms_mean": 40.070794317649586, "total_s": 3.0128028329927474, "peak_mem_gb": 2.12796398, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.7270731940070617, "cache_misses": 2860, "cache_hits": 7619, "stall_s": 1.2197487407829612, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 6221, "installed": 6261, "late": 486, "dup": 11019, "predictor": "transition2", "m": 16}, "ssd_bytes": 32137150464, "ssd_read_s": 5.259236756595783}101{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:55:42", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 425.2993749687448, "prompt_tps": 56.48295363041655, "gen_tps": 32.138579546005204, "decode_tps_wall": 31.636594953866712, "itl_ms_p50": 30.74816707521677, "itl_ms_p95": 42.22845798358321, "itl_ms_p99": 44.47245807386935, "itl_ms_mean": 31.608964285133258, "total_s": 2.41666412493214, "peak_mem_gb": 2.00596628, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.7669347805019866, "cache_misses": 2405, "cache_hits": 7914, "stall_s": 0.9113926404388621, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 4672, "installed": 4705, "late": 304, "dup": 12520, "predictor": "transition2", "m": 16}, "ssd_bytes": 25045106688, "ssd_read_s": 4.0977403610013425}102{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:55:44", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 346.1302499054, "prompt_tps": 54.953424612251396, "gen_tps": 33.33982331112324, "decode_tps_wall": 32.81931171593335, "itl_ms_p50": 29.924332979135215, "itl_ms_p95": 39.25083298236132, "itl_ms_p99": 42.250624974258244, "itl_ms_mean": 30.469865079910036, "total_s": 2.2657317499397323, "peak_mem_gb": 1.804486556, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.798721071863581, "cache_misses": 1983, "cache_hits": 7869, "stall_s": 0.8119838304119185, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 4591, "installed": 4557, "late": 371, "dup": 12153, "predictor": "transition2", "m": 16}, "ssd_bytes": 23265017856, "ssd_read_s": 3.6946225438732654}103{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:55:47", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 476.76983301062137, "prompt_tps": 60.87839121917163, "gen_tps": 27.02849420181791, "decode_tps_wall": 26.606332233733795, "itl_ms_p50": 37.03683300409466, "itl_ms_p95": 47.84154205117375, "itl_ms_p99": 52.01483308337629, "itl_ms_mean": 37.585037697609216, "total_s": 2.844627207960002, "peak_mem_gb": 2.019850032, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.7278366256322168, "cache_misses": 2852, "cache_hits": 7627, "stall_s": 1.1982156806625426, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 6246, "installed": 6233, "late": 507, "dup": 10994, "predictor": "transition2", "m": 16}, "ssd_bytes": 32197312512, "ssd_read_s": 5.143610044615343}104{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:55:50", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 368.6952500138432, "prompt_tps": 51.58245006719815, "gen_tps": 32.463807034247445, "decode_tps_wall": 31.95682615395422, "itl_ms_p50": 30.910708010196686, "itl_ms_p95": 40.77341593801975, "itl_ms_p99": 43.03674993570894, "itl_ms_mean": 31.292218920065178, "total_s": 2.3401050419779494, "peak_mem_gb": 1.968876252, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.7978075517661388, "cache_misses": 1992, "cache_hits": 7860, "stall_s": 0.8308984115719795, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 4606, "installed": 4615, "late": 358, "dup": 12138, "predictor": "transition2", "m": 16}, "ssd_bytes": 23349952512, "ssd_read_s": 3.80275591800455}105{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:55:52", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 410.11187504045665, "prompt_tps": 43.93609203570996, "gen_tps": 26.751296903221434, "decode_tps_wall": 26.33350143768132, "itl_ms_p50": 36.49824997410178, "itl_ms_p95": 50.55545805953443, "itl_ms_p99": 55.68374996073544, "itl_ms_mean": 37.97444112650636, "total_s": 2.8025016660103574, "peak_mem_gb": 1.95638236, "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": 0.7444265916850773, "cache_misses": 2545, "cache_hits": 7413, "stall_s": 1.1414121487177908, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 6657, "installed": 6643, "late": 710, "dup": 10183, "predictor": "transition2", "m": 16}, "ssd_bytes": 32565362688, "ssd_read_s": 4.980593756656162}106{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:55:55", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 646.0756659507751, "prompt_tps": 37.1689855004326, "gen_tps": 26.376589670662668, "decode_tps_wall": 25.96460213401613, "itl_ms_p50": 38.04433310870081, "itl_ms_p95": 56.77058396395296, "itl_ms_p99": 65.424875007011, "itl_ms_mean": 38.51397355671026, "total_s": 3.0724560000235215, "peak_mem_gb": 2.026643728, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.7680007752689214, "cache_misses": 2394, "cache_hits": 7925, "stall_s": 1.486848161672242, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 4463, "installed": 4517, "late": 405, "dup": 12729, "predictor": "transition2", "m": 16}, "ssd_bytes": 24266539008, "ssd_read_s": 6.124055169057101}107{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:55:59", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 378.0507919145748, "prompt_tps": 37.06682700288372, "gen_tps": 23.243584398691333, "decode_tps_wall": 22.880552968713843, "itl_ms_p50": 43.70020900387317, "itl_ms_p95": 53.16062504425645, "itl_ms_p99": 66.26566592603922, "itl_ms_mean": 43.70523742880554, "total_s": 3.131480749929324, "peak_mem_gb": 1.798916584, "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": 0.703205791106515, "cache_misses": 2870, "cache_hits": 6800, "stall_s": 1.3516874810447916, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 8046, "installed": 7980, "late": 742, "dup": 8490, "predictor": "transition2", "m": 16}, "ssd_bytes": 38631112704, "ssd_read_s": 5.961565906996839}108{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:56:02", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 595.8154579857364, "prompt_tps": 53.74817590743959, "gen_tps": 25.37077504697424, "decode_tps_wall": 24.97453118122204, "itl_ms_p50": 40.18462507519871, "itl_ms_p95": 47.956916969269514, "itl_ms_p99": 67.65291700139642, "itl_ms_mean": 40.040791666667374, "total_s": 3.118385332985781, "peak_mem_gb": 2.294852752, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.6883243338939542, "cache_misses": 3521, "cache_hits": 7776, "stall_s": 1.3618296253262088, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 6736, "installed": 6772, "late": 653, "dup": 11400, "predictor": "transition2", "m": 16}, "ssd_bytes": 36298948608, "ssd_read_s": 5.666648270911537}109{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:56:07", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 478.6368330242112, "prompt_tps": 67.70281952994452, "gen_tps": 46.02429908057159, "decode_tps_wall": 45.30772847460189, "itl_ms_p50": 21.44587505608797, "itl_ms_p95": 29.769583023153245, "itl_ms_p99": 33.46587496344, "itl_ms_mean": 22.07128968208082, "total_s": 1.869128082995303, "peak_mem_gb": 2.805416836, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.71559902200489, "cache_misses": 2908, "cache_hits": 7317, "stall_s": 0.9139544410863891, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 10291249152, "ssd_read_s": 2.1015102174133062}110{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:56:09", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 245.16537506133318, "prompt_tps": 98.05817563791334, "gen_tps": 57.37723932649922, "decode_tps_wall": 56.48113558477519, "itl_ms_p50": 17.299875034950674, "itl_ms_p95": 23.77570909447968, "itl_ms_p99": 26.005167048424482, "itl_ms_mean": 17.70502645965843, "total_s": 1.3605820420198143, "peak_mem_gb": 2.480622872, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.8336809176225235, "cache_misses": 1595, "cache_hits": 7995, "stall_s": 0.4802855326561257, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 5644615680, "ssd_read_s": 0.864837720640935}111{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:56:10", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 212.01699995435774, "prompt_tps": 66.18112848361841, "gen_tps": 44.76336658404608, "decode_tps_wall": 44.064319079396974, "itl_ms_p50": 22.347333026118577, "itl_ms_p95": 32.07187505904585, "itl_ms_p99": 34.36987497843802, "itl_ms_mean": 22.694098556207287, "total_s": 1.6417452089954168, "peak_mem_gb": 2.385389948, "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": 0.7323761200475009, "cache_misses": 2479, "cache_hits": 6784, "stall_s": 0.7234980964567512, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 8773042176, "ssd_read_s": 1.447061560349539}112{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:56:12", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 251.57787499483675, "prompt_tps": 71.65770317737694, "gen_tps": 48.42080705339218, "decode_tps_wall": 47.664664722850894, "itl_ms_p50": 20.775957964360714, "itl_ms_p95": 26.101749972440302, "itl_ms_p99": 30.63241601921618, "itl_ms_mean": 20.979902110180802, "total_s": 1.5733117079362273, "peak_mem_gb": 2.472523792, "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": 0.7912429076116048, "cache_misses": 1950, "cache_hits": 7391, "stall_s": 0.6353897843509912, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 6900940800, "ssd_read_s": 1.1909447448560968}113{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:56:13", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 289.09962507896125, "prompt_tps": 100.43671230577584, "gen_tps": 49.18196558897762, "decode_tps_wall": 48.41389732634681, "itl_ms_p50": 20.339249982498586, "itl_ms_p95": 26.802208041772246, "itl_ms_p99": 34.39366596285254, "itl_ms_mean": 20.655226189687493, "total_s": 1.5903788750292733, "peak_mem_gb": 2.597377892, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.7793997550020416, "cache_misses": 2161, "cache_hits": 7635, "stall_s": 0.6567125713918358, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 7647657984, "ssd_read_s": 1.286987041006796}114{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:56:15", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 199.78074997197837, "prompt_tps": 95.28834680741699, "gen_tps": 56.62403522263462, "decode_tps_wall": 55.73988672118333, "itl_ms_p50": 17.17858307529241, "itl_ms_p95": 24.96608393266797, "itl_ms_p99": 26.742333080619574, "itl_ms_mean": 17.940474206596495, "total_s": 1.3300306249875575, "peak_mem_gb": 2.469248884, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.8563977006599958, "cache_misses": 1349, "cache_hits": 8045, "stall_s": 0.4672505249036476, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 4774035456, "ssd_read_s": 0.7979402884375304}115{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:56:16", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 284.16120901238173, "prompt_tps": 112.75508099171603, "gen_tps": 44.10160803690054, "decode_tps_wall": 43.41288564870694, "itl_ms_p50": 22.570249973796308, "itl_ms_p95": 30.478207976557314, "itl_ms_p99": 33.801999990828335, "itl_ms_mean": 23.034635570920294, "total_s": 1.7353432499803603, "peak_mem_gb": 2.755937396, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.7463080684596577, "cache_misses": 2594, "cache_hits": 7631, "stall_s": 0.7904208408435807, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 9180020736, "ssd_read_s": 1.5752222624141723}116{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:56:18", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 238.3841669652611, "prompt_tps": 75.60417725501566, "gen_tps": 51.33413095018941, "decode_tps_wall": 50.53244724218592, "itl_ms_p50": 19.192834035493433, "itl_ms_p95": 24.936707923188806, "itl_ms_p99": 27.704833075404167, "itl_ms_mean": 19.78926520632018, "total_s": 1.4851078749634326, "peak_mem_gb": 2.464595216, "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": 0.7930628412375549, "cache_misses": 1933, "cache_hits": 7408, "stall_s": 0.5997709431685507, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 6840778752, "ssd_read_s": 1.1616856374312192}117{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:56:20", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 221.29808401223272, "prompt_tps": 63.352071222316766, "gen_tps": 44.45762609825912, "decode_tps_wall": 43.763362020621706, "itl_ms_p50": 22.22233300562948, "itl_ms_p95": 30.22420802153647, "itl_ms_p99": 33.73758401721716, "itl_ms_mean": 22.850164014565213, "total_s": 1.660858416929841, "peak_mem_gb": 2.374838524, "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": 0.7287056029364137, "cache_misses": 2513, "cache_hits": 6750, "stall_s": 0.744502259651199, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 8893366272, "ssd_read_s": 1.5194422950735316}118{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:56:21", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 269.4497909396887, "prompt_tps": 107.76387631799321, "gen_tps": 49.1663455435908, "decode_tps_wall": 48.398579948683256, "itl_ms_p50": 20.051375031471252, "itl_ms_p95": 28.881374979391694, "itl_ms_p99": 36.583874956704676, "itl_ms_mean": 20.661763238927556, "total_s": 1.5711408749921247, "peak_mem_gb": 2.607646564, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.7755206206614945, "cache_misses": 2199, "cache_hits": 7597, "stall_s": 0.6573501937091351, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 7782137856, "ssd_read_s": 1.309832967002876}119{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:56:23", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 235.84695789031684, "prompt_tps": 101.89591833638708, "gen_tps": 56.34368446101921, "decode_tps_wall": 55.463920645825475, "itl_ms_p50": 17.546750023029745, "itl_ms_p95": 24.633875000290573, "itl_ms_p99": 26.978000067174435, "itl_ms_mean": 18.029738762711602, "total_s": 1.3717204999411479, "peak_mem_gb": 2.480622872, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.8375391032325339, "cache_misses": 1558, "cache_hits": 8032, "stall_s": 0.4964398378506303, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 5513674752, "ssd_read_s": 0.9026410088408738}120{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:56:24", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 190.15433301683515, "prompt_tps": 100.08026174828318, "gen_tps": 58.397335353996795, "decode_tps_wall": 57.48555662623313, "itl_ms_p50": 16.422374988906085, "itl_ms_p95": 23.83612492121756, "itl_ms_p99": 25.87599994149059, "itl_ms_mean": 17.39567395166627, "total_s": 1.2860817919718102, "peak_mem_gb": 2.469347188, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.8592718756653183, "cache_misses": 1322, "cache_hits": 8072, "stall_s": 0.44590369798243046, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 4678483968, "ssd_read_s": 0.7657993045868352}121{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:56:25", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 274.3655410595238, "prompt_tps": 105.82126424171258, "gen_tps": 49.77887291230627, "decode_tps_wall": 49.00150359764567, "itl_ms_p50": 20.00466699246317, "itl_ms_p95": 26.791749987751245, "itl_ms_p99": 34.54133297782391, "itl_ms_mean": 20.40753704643557, "total_s": 1.5600403749849647, "peak_mem_gb": 2.57641866, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.7832788893425888, "cache_misses": 2123, "cache_hits": 7673, "stall_s": 0.6371748555684462, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 7513178112, "ssd_read_s": 1.2614294610684738}122{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:56:27", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 198.3139580115676, "prompt_tps": 95.95129713290473, "gen_tps": 57.90602971523838, "decode_tps_wall": 57.00180675362896, "itl_ms_p50": 16.657334053888917, "itl_ms_p95": 24.488208000548184, "itl_ms_p99": 25.874500046484172, "itl_ms_mean": 17.543303571449268, "total_s": 1.3035420830128714, "peak_mem_gb": 2.469248884, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.8563977006599958, "cache_misses": 1349, "cache_hits": 8045, "stall_s": 0.4601672978606075, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 4774035456, "ssd_read_s": 0.802551290136762}123{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:56:28", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 226.38579201884568, "prompt_tps": 79.614709048029, "gen_tps": 52.61680569524645, "decode_tps_wall": 51.79508333273456, "itl_ms_p50": 19.09966696985066, "itl_ms_p95": 24.101708899252117, "itl_ms_p99": 27.171374997124076, "itl_ms_mean": 19.306851841051074, "total_s": 1.4427174580050632, "peak_mem_gb": 2.453978384, "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": 0.7957392142168933, "cache_misses": 1908, "cache_hits": 7433, "stall_s": 0.5772590979468077, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 6752305152, "ssd_read_s": 1.1212810397846624}124{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:56:30", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 236.9219169486314, "prompt_tps": 101.43647058367213, "gen_tps": 57.36168942730434, "decode_tps_wall": 56.4659760563488, "itl_ms_p50": 17.01224991120398, "itl_ms_p95": 24.26766697317362, "itl_ms_p99": 26.685958029702306, "itl_ms_mean": 17.7097797619238, "total_s": 1.352638041949831, "peak_mem_gb": 2.48770076, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.8338894681960375, "cache_misses": 1593, "cache_hits": 7997, "stall_s": 0.4866465798113495, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 5637537792, "ssd_read_s": 0.898337729042396}125{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:56:31", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 209.65387509204447, "prompt_tps": 66.86458074667341, "gen_tps": 45.7667499358963, "decode_tps_wall": 45.052025726211426, "itl_ms_p50": 21.982959005981684, "itl_ms_p95": 28.87145895510912, "itl_ms_p99": 31.690999981947243, "itl_ms_mean": 22.19656017416763, "total_s": 1.6080371660646051, "peak_mem_gb": 2.36304204, "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": 0.7323761200475009, "cache_misses": 2479, "cache_hits": 6784, "stall_s": 0.7163165680831298, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 8773042176, "ssd_read_s": 1.480862574884668}126{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2574612"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:56:33", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 273.5227909870446, "prompt_tps": 117.1461191764391, "gen_tps": 47.4691289863635, "decode_tps_wall": 46.72785852638877, "itl_ms_p50": 20.936249988153577, "itl_ms_p95": 27.268707985058427, "itl_ms_p99": 31.281375093385577, "itl_ms_mean": 21.40050992140517, "total_s": 1.6217549160355702, "peak_mem_gb": 2.751677372, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.7433740831295843, "cache_misses": 2624, "cache_hits": 7601, "stall_s": 0.7309220866300166, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 9286189056, "ssd_read_s": 1.5045044647995383}127{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:56:40", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 537.12837502826, "prompt_tps": 60.01188509066157, "gen_tps": 38.37134619658702, "decode_tps_wall": 37.77377745861214, "itl_ms_p50": 26.52791701257229, "itl_ms_p95": 31.685000052675605, "itl_ms_p99": 41.86441597994417, "itl_ms_mean": 26.47339152393951, "total_s": 2.2049520410364494, "peak_mem_gb": 2.8607033, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.7713447432762837, "cache_misses": 2338, "cache_hits": 7887, "stall_s": 0.896944917505607, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1456, "installed": 1456, "late": 32, "dup": 6824, "predictor": "transition", "m": 8}, "ssd_bytes": 13426753536, "ssd_read_s": 2.999714838573709}128{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:56:42", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 248.5780839342624, "prompt_tps": 96.6808910407994, "gen_tps": 47.010754168751824, "decode_tps_wall": 46.27650006579437, "itl_ms_p50": 21.070166025310755, "itl_ms_p95": 28.713042032904923, "itl_ms_p99": 30.693084001541138, "itl_ms_mean": 21.6092400803482, "total_s": 1.609960208996199, "peak_mem_gb": 2.487737752, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.8579770594369135, "cache_misses": 1362, "cache_hits": 8228, "stall_s": 0.46157957590185106, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 397, "installed": 397, "late": 14, "dup": 7755, "predictor": "transition", "m": 8}, "ssd_bytes": 6225002496, "ssd_read_s": 1.0489242697367445}129{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:56:44", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 216.25570906326175, "prompt_tps": 64.83709027208769, "gen_tps": 35.52815117461518, "decode_tps_wall": 34.97319774721887, "itl_ms_p50": 28.280625003390014, "itl_ms_p95": 36.04683291632682, "itl_ms_p99": 41.89708305057138, "itl_ms_mean": 28.593324729064037, "total_s": 2.017635166994296, "peak_mem_gb": 2.377201144, "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": 0.7943430853935011, "cache_misses": 1905, "cache_hits": 7358, "stall_s": 0.733634035452269, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2211, "installed": 2207, "late": 68, "dup": 5837, "predictor": "transition", "m": 8}, "ssd_bytes": 14566293504, "ssd_read_s": 2.5604033486451954}130{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:56:46", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 234.04616699554026, "prompt_tps": 77.01186831209532, "gen_tps": 42.0842544781391, "decode_tps_wall": 41.42692295140082, "itl_ms_p50": 23.884625057689846, "itl_ms_p95": 29.785625054500997, "itl_ms_p99": 34.03600002638996, "itl_ms_mean": 24.13889154097036, "total_s": 1.7547963340766728, "peak_mem_gb": 2.50075918, "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": 0.8352424793919281, "cache_misses": 1539, "cache_hits": 7802, "stall_s": 0.5556992036290467, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1089, "installed": 1092, "late": 27, "dup": 6983, "predictor": "transition", "m": 8}, "ssd_bytes": 9300344832, "ssd_read_s": 1.6137432872783393}131{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:56:48", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 281.386791029945, "prompt_tps": 103.17887315821685, "gen_tps": 33.53500694342829, "decode_tps_wall": 33.01123075692832, "itl_ms_p50": 29.71879206597805, "itl_ms_p95": 39.29099999368191, "itl_ms_p99": 41.45362495910376, "itl_ms_mean": 30.292720903479864, "total_s": 2.1898282079491764, "peak_mem_gb": 2.594692576, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.81472029399755, "cache_misses": 1815, "cache_hits": 7981, "stall_s": 0.6964973034337163, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 952, "installed": 950, "late": 21, "dup": 7224, "predictor": "transition", "m": 8}, "ssd_bytes": 9792258048, "ssd_read_s": 1.8330684097018093}132{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:56:49", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 234.94666605256498, "prompt_tps": 81.0111470370424, "gen_tps": 44.660945356483474, "decode_tps_wall": 43.96355397279772, "itl_ms_p50": 21.836041007190943, "itl_ms_p95": 29.456667019985616, "itl_ms_p99": 33.34979200735688, "itl_ms_mean": 22.746113760929024, "total_s": 1.6679518329910934, "peak_mem_gb": 2.476345452, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.8776878858846072, "cache_misses": 1149, "cache_hits": 8245, "stall_s": 0.4518067493336275, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 331, "installed": 334, "late": 2, "dup": 7765, "predictor": "transition", "m": 8}, "ssd_bytes": 5237637120, "ssd_read_s": 0.9524213743861765}133{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:56:52", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 336.71350008808076, "prompt_tps": 95.1377192722343, "gen_tps": 31.770734437612802, "decode_tps_wall": 31.27466280226716, "itl_ms_p50": 31.3452499685809, "itl_ms_p95": 39.24054198432714, "itl_ms_p99": 50.376374972984195, "itl_ms_mean": 31.97476520602192, "total_s": 2.351123708067462, "peak_mem_gb": 2.769020348, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.7899266503667481, "cache_misses": 2148, "cache_hits": 8077, "stall_s": 0.8536091272253543, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1285, "installed": 1285, "late": 28, "dup": 6995, "predictor": "transition", "m": 8}, "ssd_bytes": 12149194752, "ssd_read_s": 2.438564954791218}134{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:56:54", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 265.23241598624736, "prompt_tps": 68.02827140271641, "gen_tps": 35.15052388005113, "decode_tps_wall": 34.60156932921162, "itl_ms_p50": 28.241583961062133, "itl_ms_p95": 37.11599996313453, "itl_ms_p99": 44.96050009038299, "itl_ms_mean": 28.900423286748783, "total_s": 2.0859590830514207, "peak_mem_gb": 2.463500932, "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": 0.8331013810084573, "cache_misses": 1559, "cache_hits": 7782, "stall_s": 0.6608698770869523, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1048, "installed": 1047, "late": 32, "dup": 7024, "predictor": "transition", "m": 8}, "ssd_bytes": 9226027008, "ssd_read_s": 1.84350956894923}135{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:56:56", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 239.99174998607486, "prompt_tps": 58.42973215935761, "gen_tps": 33.53723287921878, "decode_tps_wall": 33.0134125450079, "itl_ms_p50": 30.094875022768974, "itl_ms_p95": 38.700582925230265, "itl_ms_p99": 46.733499970287085, "itl_ms_mean": 30.29071892027758, "total_s": 2.1483070419635624, "peak_mem_gb": 2.382968316, "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": 0.7921839576810968, "cache_misses": 1925, "cache_hits": 7338, "stall_s": 0.7835988189326599, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2218, "installed": 2215, "late": 75, "dup": 5830, "predictor": "transition", "m": 8}, "ssd_bytes": 14661844992, "ssd_read_s": 2.7195660427678376}136{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:56:58", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 279.3260000180453, "prompt_tps": 103.95477738535655, "gen_tps": 42.96203901905487, "decode_tps_wall": 42.29114396375833, "itl_ms_p50": 24.486916954629123, "itl_ms_p95": 30.314750038087368, "itl_ms_p99": 32.26291702594608, "itl_ms_mean": 23.64561244446252, "total_s": 1.7689995840191841, "peak_mem_gb": 2.625694176, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.8138015516537362, "cache_misses": 1824, "cache_hits": 7972, "stall_s": 0.6315422409679741, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 978, "installed": 979, "late": 28, "dup": 7198, "predictor": "transition", "m": 8}, "ssd_bytes": 9916121088, "ssd_read_s": 1.745780982542783}137{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:56:59", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 246.90954200923443, "prompt_tps": 97.33786021979385, "gen_tps": 47.21045821123445, "decode_tps_wall": 46.47319188429845, "itl_ms_p50": 21.33366698399186, "itl_ms_p95": 27.76787499897182, "itl_ms_p99": 31.163415987975895, "itl_ms_mean": 21.517781745864166, "total_s": 1.602529791998677, "peak_mem_gb": 2.487737752, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.8604796663190823, "cache_misses": 1338, "cache_hits": 8252, "stall_s": 0.4553022999316454, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 399, "installed": 402, "late": 10, "dup": 7753, "predictor": "transition", "m": 8}, "ssd_bytes": 6147145728, "ssd_read_s": 1.0380624668905511}138{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:57:01", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 200.95729094464332, "prompt_tps": 94.68678367729349, "gen_tps": 47.840757566434014, "decode_tps_wall": 47.093690131907024, "itl_ms_p50": 20.65054199192673, "itl_ms_p95": 28.029291075654328, "itl_ms_p99": 32.08879102021456, "itl_ms_mean": 21.234267206478215, "total_s": 1.5387161249527708, "peak_mem_gb": 2.47646014, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.8797104534809452, "cache_misses": 1130, "cache_hits": 8264, "stall_s": 0.4055240408051759, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 325, "installed": 325, "late": 3, "dup": 7771, "predictor": "transition", "m": 8}, "ssd_bytes": 5149163520, "ssd_read_s": 0.8579142205417156}139{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:57:03", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 282.0582080166787, "prompt_tps": 102.92806675704358, "gen_tps": 40.68185545852985, "decode_tps_wall": 40.04648252502245, "itl_ms_p50": 24.982333066873252, "itl_ms_p95": 30.436999979428947, "itl_ms_p99": 33.80570805165917, "itl_ms_mean": 24.970982142443223, "total_s": 1.8552300829906017, "peak_mem_gb": 2.575310304, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.8164556962025317, "cache_misses": 1798, "cache_hits": 7998, "stall_s": 0.6233157982351258, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 935, "installed": 932, "late": 27, "dup": 7241, "predictor": "transition", "m": 8}, "ssd_bytes": 9671933952, "ssd_read_s": 1.675455072778277}140{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:57:04", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 204.97587497811764, "prompt_tps": 92.85052972381513, "gen_tps": 47.64513307839783, "decode_tps_wall": 46.901075079288255, "itl_ms_p50": 20.554249989800155, "itl_ms_p95": 28.412916930392385, "itl_ms_p99": 29.660750064067543, "itl_ms_mean": 21.32147287262515, "total_s": 1.548228665953502, "peak_mem_gb": 2.479884396, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.8776878858846072, "cache_misses": 1149, "cache_hits": 8245, "stall_s": 0.41758362122345716, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 332, "installed": 335, "late": 4, "dup": 7764, "predictor": "transition", "m": 8}, "ssd_bytes": 5241176064, "ssd_read_s": 0.9039581266697496}141{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:57:06", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 229.15124997962266, "prompt_tps": 78.66005497161797, "gen_tps": 41.96813046903063, "decode_tps_wall": 41.31264592547789, "itl_ms_p50": 24.0199159597978, "itl_ms_p95": 30.005750013515353, "itl_ms_p99": 33.73841708526015, "itl_ms_mean": 24.20566336525279, "total_s": 1.7541080419905484, "peak_mem_gb": 2.450329484, "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": 0.8359918638261428, "cache_misses": 1532, "cache_hits": 7809, "stall_s": 0.5551461785798892, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1046, "installed": 1046, "late": 31, "dup": 7026, "predictor": "transition", "m": 8}, "ssd_bytes": 9123397632, "ssd_read_s": 1.5742781347362325}142{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:57:08", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 250.88149996008724, "prompt_tps": 95.79490313426386, "gen_tps": 46.538314452759366, "decode_tps_wall": 45.81149890277602, "itl_ms_p50": 21.1615419248119, "itl_ms_p95": 28.868083958514035, "itl_ms_p99": 31.803624937310815, "itl_ms_mean": 21.82858068281637, "total_s": 1.6260820829775184, "peak_mem_gb": 2.49481564, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.8584984358706986, "cache_misses": 1357, "cache_hits": 8233, "stall_s": 0.4639157010242343, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 411, "installed": 411, "late": 12, "dup": 7741, "predictor": "transition", "m": 8}, "ssd_bytes": 6256852992, "ssd_read_s": 1.0651860805228353}143{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:57:10", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 216.87979099806398, "prompt_tps": 64.6351479425608, "gen_tps": 35.75213519549105, "decode_tps_wall": 35.19379641972889, "itl_ms_p50": 28.78604200668633, "itl_ms_p95": 36.30054194945842, "itl_ms_p99": 41.629457962699234, "itl_ms_mean": 28.41409855514824, "total_s": 2.006967999972403, "peak_mem_gb": 2.384213624, "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": 0.7944510417791212, "cache_misses": 1904, "cache_hits": 7359, "stall_s": 0.7393082856433466, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2211, "installed": 2208, "late": 75, "dup": 5837, "predictor": "transition", "m": 8}, "ssd_bytes": 14562754560, "ssd_read_s": 2.5910886902129278}144{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:57:12", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 317.04166601412, "prompt_tps": 101.03133513549277, "gen_tps": 38.2362503826078, "decode_tps_wall": 37.63909005929106, "itl_ms_p50": 26.37291606515646, "itl_ms_p95": 31.681708991527557, "itl_ms_p99": 40.97141698002815, "itl_ms_mean": 26.568123682712514, "total_s": 1.9908334580250084, "peak_mem_gb": 2.802509684, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.7899266503667481, "cache_misses": 2148, "cache_hits": 8077, "stall_s": 0.7414987228112295, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1326, "installed": 1329, "late": 26, "dup": 6954, "predictor": "transition", "m": 8}, "ssd_bytes": 12294291456, "ssd_read_s": 2.2050836889538914}145{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:57:19", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 526.2467919383198, "prompt_tps": 61.317796123046456, "gen_tps": 38.66756407637023, "decode_tps_wall": 38.06518490863805, "itl_ms_p50": 26.244333013892174, "itl_ms_p95": 31.38954099267721, "itl_ms_p99": 42.13595809414983, "itl_ms_mean": 26.270724873664598, "total_s": 2.1813024589791894, "peak_mem_gb": 2.837896772, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.7598044009779951, "cache_misses": 2456, "cache_hits": 7769, "stall_s": 0.8876701433910057, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1171, "installed": 1171, "late": 9, "dup": 7077, "predictor": "transition2", "m": 8}, "ssd_bytes": 12835749888, "ssd_read_s": 2.7585903315339237}146{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:57:21", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 246.94054201245308, "prompt_tps": 97.32212104346677, "gen_tps": 47.01821700718641, "decode_tps_wall": 46.283937078079965, "itl_ms_p50": 21.34562493301928, "itl_ms_p95": 28.52187503594905, "itl_ms_p99": 30.362334102392197, "itl_ms_mean": 21.605767856632905, "total_s": 1.6081039169803262, "peak_mem_gb": 2.491276696, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.8531803962460897, "cache_misses": 1408, "cache_hits": 8182, "stall_s": 0.4605353972874582, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 340, "installed": 340, "late": 4, "dup": 7812, "predictor": "transition2", "m": 8}, "ssd_bytes": 6186074112, "ssd_read_s": 1.021657341858372}147{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:57:23", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 220.02695803530514, "prompt_tps": 63.71451033254018, "gen_tps": 35.766929823448926, "decode_tps_wall": 35.2083281655901, "itl_ms_p50": 28.530958923511207, "itl_ms_p95": 37.50645800027996, "itl_ms_p99": 42.41429199464619, "itl_ms_mean": 28.402371032695687, "total_s": 2.0093763330951333, "peak_mem_gb": 2.377201272, "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": 0.7828997085177588, "cache_misses": 2011, "cache_hits": 7252, "stall_s": 0.7473375742556527, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2012, "installed": 2007, "late": 55, "dup": 6036, "predictor": "transition2", "m": 8}, "ssd_bytes": 14237171712, "ssd_read_s": 2.4921520388452336}148{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:57:24", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 239.7110000019893, "prompt_tps": 75.1862686549162, "gen_tps": 42.304296297558295, "decode_tps_wall": 41.64360362790369, "itl_ms_p50": 24.068666039966047, "itl_ms_p95": 30.433833017013967, "itl_ms_p99": 32.19966706819832, "itl_ms_mean": 24.01329166743727, "total_s": 1.7525483750505373, "peak_mem_gb": 2.479410828, "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": 0.8257145915854833, "cache_misses": 1628, "cache_hits": 7713, "stall_s": 0.561451684916392, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 930, "installed": 934, "late": 32, "dup": 7142, "predictor": "transition2", "m": 8}, "ssd_bytes": 9052618752, "ssd_read_s": 1.5275575995910913}149{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:57:26", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 264.22424998600036, "prompt_tps": 109.88906891185485, "gen_tps": 42.673507455874024, "decode_tps_wall": 42.00696497735514, "itl_ms_p50": 24.298166972585022, "itl_ms_p95": 29.134584008716047, "itl_ms_p99": 32.61645801831037, "itl_ms_mean": 23.805576064328235, "total_s": 1.7639755420386791, "peak_mem_gb": 2.604460512, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.8126786443446304, "cache_misses": 1835, "cache_hits": 7961, "stall_s": 0.6109398454427719, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 802, "installed": 803, "late": 22, "dup": 7374, "predictor": "transition2", "m": 8}, "ssd_bytes": 9332195328, "ssd_read_s": 1.5725603667087853}150{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:57:28", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 204.57962492946535, "prompt_tps": 93.01282012853194, "gen_tps": 49.11216022914396, "decode_tps_wall": 48.345155234164295, "itl_ms_p50": 20.258666947484016, "itl_ms_p95": 26.490667019970715, "itl_ms_p99": 29.32937501464039, "itl_ms_mean": 20.68459590534783, "total_s": 1.5077091669663787, "peak_mem_gb": 2.472921196, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.8742814562486694, "cache_misses": 1181, "cache_hits": 8213, "stall_s": 0.4042675201781094, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 240, "installed": 240, "late": 5, "dup": 7856, "predictor": "transition2", "m": 8}, "ssd_bytes": 5028839424, "ssd_read_s": 0.8246123712742701}151{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:57:30", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 298.8200420513749, "prompt_tps": 107.19331348912667, "gen_tps": 38.62533536907778, "decode_tps_wall": 38.022065019358564, "itl_ms_p50": 26.01483406033367, "itl_ms_p95": 31.988083035685122, "itl_ms_p99": 39.039624971337616, "itl_ms_mean": 26.300517856956475, "total_s": 1.9557526670396328, "peak_mem_gb": 2.769020532, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.7831784841075795, "cache_misses": 2217, "cache_hits": 8008, "stall_s": 0.7259550300659612, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1062, "installed": 1062, "late": 10, "dup": 7218, "predictor": "transition2", "m": 8}, "ssd_bytes": 11604197376, "ssd_read_s": 2.00498158717528}152{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:57:31", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 235.45412498060614, "prompt_tps": 76.54730555019555, "gen_tps": 42.16143564477407, "decode_tps_wall": 41.502934349015625, "itl_ms_p50": 23.894666926935315, "itl_ms_p95": 30.284166103228927, "itl_ms_p99": 31.48029197473079, "itl_ms_mean": 24.094681874553245, "total_s": 1.7534190830774605, "peak_mem_gb": 2.468040332, "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": 0.8249652071512686, "cache_misses": 1635, "cache_hits": 7706, "stall_s": 0.5660235554678366, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 909, "installed": 908, "late": 29, "dup": 7163, "predictor": "transition2", "m": 8}, "ssd_bytes": 9003073536, "ssd_read_s": 1.5280321226455271}153{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:57:33", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 221.29662509541959, "prompt_tps": 63.34772376805383, "gen_tps": 35.47670939707307, "decode_tps_wall": 34.92266006485693, "itl_ms_p50": 28.77699991222471, "itl_ms_p95": 36.66425007395446, "itl_ms_p99": 43.07016695383936, "itl_ms_mean": 28.634703030721063, "total_s": 2.0252829160308465, "peak_mem_gb": 2.381919772, "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": 0.7798769297203929, "cache_misses": 2039, "cache_hits": 7224, "stall_s": 0.7604859068524092, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2011, "installed": 2007, "late": 66, "dup": 6037, "predictor": "transition2", "m": 8}, "ssd_bytes": 14332723200, "ssd_read_s": 2.531662316294387}154{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:57:35", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 277.9518330935389, "prompt_tps": 104.46504005461514, "gen_tps": 41.49031706884608, "decode_tps_wall": 40.842303347646805, "itl_ms_p50": 24.697249988093972, "itl_ms_p95": 29.536165995523334, "itl_ms_p99": 32.76883403304964, "itl_ms_mean": 24.484417332883275, "total_s": 1.8204701250651851, "peak_mem_gb": 2.630277664, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.810432829726419, "cache_misses": 1857, "cache_hits": 7939, "stall_s": 0.6138930609449744, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 810, "installed": 815, "late": 24, "dup": 7366, "predictor": "transition2", "m": 8}, "ssd_bytes": 9438363648, "ssd_read_s": 1.5867449769284576}155{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:57:37", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 242.76633304543793, "prompt_tps": 98.9883257059173, "gen_tps": 46.73172294766079, "decode_tps_wall": 46.00191204773223, "itl_ms_p50": 21.043957909569144, "itl_ms_p95": 29.10587494261563, "itl_ms_p99": 30.490832985378802, "itl_ms_mean": 21.73822685810072, "total_s": 1.6122746251057833, "peak_mem_gb": 2.487737752, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.8558915537017727, "cache_misses": 1382, "cache_hits": 8208, "stall_s": 0.45742425217758864, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 337, "installed": 337, "late": 5, "dup": 7815, "predictor": "transition2", "m": 8}, "ssd_bytes": 6083444736, "ssd_read_s": 1.007557698409073}156{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:57:38", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 202.49391696415842, "prompt_tps": 93.97415458566383, "gen_tps": 48.16626846601871, "decode_tps_wall": 47.4140570839877, "itl_ms_p50": 20.457208971492946, "itl_ms_p95": 28.0318750301376, "itl_ms_p99": 29.554625041782856, "itl_ms_mean": 21.090791666037624, "total_s": 1.5312137919245288, "peak_mem_gb": 2.476525676, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.8752395145837769, "cache_misses": 1172, "cache_hits": 8222, "stall_s": 0.40988460066728294, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 234, "installed": 234, "late": 6, "dup": 7862, "predictor": "transition2", "m": 8}, "ssd_bytes": 4975755264, "ssd_read_s": 0.8229692546883598}157{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:57:40", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 250.01370802056044, "prompt_tps": 116.12952296055184, "gen_tps": 41.96681521643548, "decode_tps_wall": 41.311333190355455, "itl_ms_p50": 24.232625029981136, "itl_ms_p95": 30.34666704479605, "itl_ms_p99": 33.43629091978073, "itl_ms_mean": 24.206432539762723, "total_s": 1.775018958025612, "peak_mem_gb": 2.579209696, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.81472029399755, "cache_misses": 1815, "cache_hits": 7981, "stall_s": 0.6051625002874061, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 788, "installed": 788, "late": 20, "dup": 7388, "predictor": "transition2", "m": 8}, "ssd_bytes": 9211871232, "ssd_read_s": 1.5440555774839595}158{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:57:42", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 207.83304201904684, "prompt_tps": 91.55174209258396, "gen_tps": 48.629504350818316, "decode_tps_wall": 47.87007910969626, "itl_ms_p50": 20.405165967531502, "itl_ms_p95": 27.231291984207928, "itl_ms_p99": 29.89404194522649, "itl_ms_mean": 20.88987565089372, "total_s": 1.5238952080253512, "peak_mem_gb": 2.472921196, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.8742814562486694, "cache_misses": 1181, "cache_hits": 8213, "stall_s": 0.4092693250859156, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 240, "installed": 240, "late": 5, "dup": 7856, "predictor": "transition2", "m": 8}, "ssd_bytes": 5028839424, "ssd_read_s": 0.8329627900384367}159{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:57:44", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 230.12087505776435, "prompt_tps": 78.33347006973082, "gen_tps": 42.08169372605458, "decode_tps_wall": 41.42447254162209, "itl_ms_p50": 23.665833054110408, "itl_ms_p95": 30.684374971315265, "itl_ms_p99": 31.604208052158356, "itl_ms_mean": 24.14031944511133, "total_s": 1.7509610000997782, "peak_mem_gb": 2.453983116, "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": 0.8271063055347393, "cache_misses": 1615, "cache_hits": 7726, "stall_s": 0.564033625065349, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 906, "installed": 905, "late": 32, "dup": 7166, "predictor": "transition2", "m": 8}, "ssd_bytes": 8921677824, "ssd_read_s": 1.512469703797251}160{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:57:45", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 245.58254098519683, "prompt_tps": 97.85678351701762, "gen_tps": 46.77768021154988, "decode_tps_wall": 46.04708745056892, "itl_ms_p50": 21.15079201757908, "itl_ms_p95": 28.85829098522663, "itl_ms_p99": 29.13104195613414, "itl_ms_mean": 21.71690014213146, "total_s": 1.613747249939479, "peak_mem_gb": 2.491276696, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.8530761209593326, "cache_misses": 1409, "cache_hits": 8181, "stall_s": 0.45696721330750734, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 345, "installed": 346, "late": 4, "dup": 7807, "predictor": "transition2", "m": 8}, "ssd_bytes": 6207307776, "ssd_read_s": 1.0156202514190227}161{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:57:47", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 210.71124996524304, "prompt_tps": 66.53070136420489, "gen_tps": 36.08053993946833, "decode_tps_wall": 35.516990081721616, "itl_ms_p50": 28.451750054955482, "itl_ms_p95": 35.19320790655911, "itl_ms_p99": 39.47437496390194, "itl_ms_mean": 28.155539016653265, "total_s": 1.9845102080143988, "peak_mem_gb": 2.3688782, "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": 0.7826837957465184, "cache_misses": 2013, "cache_hits": 7250, "stall_s": 0.7126202502986416, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2012, "installed": 2007, "late": 46, "dup": 6036, "predictor": "transition2", "m": 8}, "ssd_bytes": 14244249600, "ssd_read_s": 2.3624545776983723}162{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:57:49", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 313.0920419935137, "prompt_tps": 102.30909516845405, "gen_tps": 38.29381507229609, "decode_tps_wall": 37.695758040872974, "itl_ms_p50": 26.349458028562367, "itl_ms_p95": 32.40716701839119, "itl_ms_p99": 42.20050002913922, "itl_ms_mean": 26.528183858664264, "total_s": 1.9843676250893623, "peak_mem_gb": 2.796152436, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.7817114914425428, "cache_misses": 2232, "cache_hits": 7993, "stall_s": 0.7305938890203834, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1084, "installed": 1089, "late": 10, "dup": 7196, "predictor": "transition2", "m": 8}, "ssd_bytes": 11735138304, "ssd_read_s": 2.0255970461294055}163{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:57:57", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 530.6949169607833, "prompt_tps": 60.711212878555976, "gen_tps": 37.61759079615561, "decode_tps_wall": 37.03165427681308, "itl_ms_p50": 26.682166033424437, "itl_ms_p95": 32.583207939751446, "itl_ms_p99": 40.29008303768933, "itl_ms_mean": 27.003924602583524, "total_s": 2.2319421669235453, "peak_mem_gb": 2.834357828, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.7527628361858191, "cache_misses": 2528, "cache_hits": 7697, "stall_s": 0.9040113523369655, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1104, "installed": 1104, "late": 10, "dup": 7104, "predictor": "transition3", "m": 8}, "ssd_bytes": 12853444608, "ssd_read_s": 2.7384828088106588}164{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:57:58", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 250.18979201558977, "prompt_tps": 96.07380080962646, "gen_tps": 47.3216806774956, "decode_tps_wall": 46.58264682391963, "itl_ms_p50": 21.394332987256348, "itl_ms_p95": 28.589375084266067, "itl_ms_p99": 29.91420798934996, "itl_ms_mean": 21.46722155527049, "total_s": 1.6026247499976307, "peak_mem_gb": 2.487737752, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.8495307612095934, "cache_misses": 1443, "cache_hits": 8147, "stall_s": 0.46130490279756486, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 277, "installed": 277, "late": 2, "dup": 7875, "predictor": "transition3", "m": 8}, "ssd_bytes": 6086983680, "ssd_read_s": 0.9874504569452256}165{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:58:00", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 218.65341695956886, "prompt_tps": 64.1226758520243, "gen_tps": 35.4168779949671, "decode_tps_wall": 34.86369104309848, "itl_ms_p50": 28.66358298342675, "itl_ms_p95": 36.35525004938245, "itl_ms_p99": 42.23512497264892, "itl_ms_mean": 28.68313623947047, "total_s": 2.0256910000462085, "peak_mem_gb": 2.363045368, "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": 0.7735075029688006, "cache_misses": 2098, "cache_hits": 7165, "stall_s": 0.7597819989314303, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1827, "installed": 1824, "late": 38, "dup": 6221, "predictor": "transition3", "m": 8}, "ssd_bytes": 13890355200, "ssd_read_s": 2.45023709803354}166{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:58:02", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 241.28454201854765, "prompt_tps": 74.71593684323653, "gen_tps": 42.30884779451468, "decode_tps_wall": 41.64811623298193, "itl_ms_p50": 23.504958022385836, "itl_ms_p95": 31.18704108055681, "itl_ms_p99": 33.52537506725639, "itl_ms_mean": 24.010689809016647, "total_s": 1.7539579999865964, "peak_mem_gb": 2.48298254, "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": 0.8210041751418478, "cache_misses": 1672, "cache_hits": 7669, "stall_s": 0.5712962452089414, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 807, "installed": 810, "late": 19, "dup": 7265, "predictor": "transition3", "m": 8}, "ssd_bytes": 8773042176, "ssd_read_s": 1.5135142977815121}167{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:58:04", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 288.3212079759687, "prompt_tps": 100.69813039139494, "gen_tps": 40.75601669777822, "decode_tps_wall": 40.11950729849327, "itl_ms_p50": 25.002750102430582, "itl_ms_p95": 31.589207937940955, "itl_ms_p99": 32.96391700860113, "itl_ms_mean": 24.925530429870363, "total_s": 1.8586296250578016, "peak_mem_gb": 2.597382624, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.8037974683544303, "cache_misses": 1922, "cache_hits": 7874, "stall_s": 0.6321207407163456, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 761, "installed": 761, "late": 15, "dup": 7415, "predictor": "transition3", "m": 8}, "ssd_bytes": 9494986752, "ssd_read_s": 1.616611955803819}168{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:58:05", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 204.7907910309732, "prompt_tps": 92.92915099835191, "gen_tps": 47.14587731748471, "decode_tps_wall": 46.4096218191413, "itl_ms_p50": 20.818791002966464, "itl_ms_p95": 28.576500015333295, "itl_ms_p99": 32.01054106466472, "itl_ms_mean": 21.547255952591225, "total_s": 1.5622679160442203, "peak_mem_gb": 2.470075244, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.8687460080902704, "cache_misses": 1233, "cache_hits": 8161, "stall_s": 0.43462368892505765, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 193, "installed": 193, "late": 1, "dup": 7903, "predictor": "transition3", "m": 8}, "ssd_bytes": 5046534144, "ssd_read_s": 0.8415085544111207}169{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:58:07", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 306.4874999690801, "prompt_tps": 104.52459468294853, "gen_tps": 37.95241014117154, "decode_tps_wall": 37.35966681644194, "itl_ms_p50": 26.526625035330653, "itl_ms_p95": 32.675832975655794, "itl_ms_p99": 37.84616698976606, "itl_ms_mean": 26.766834000775976, "total_s": 1.9927980420179665, "peak_mem_gb": 2.769020532, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.7764303178484108, "cache_misses": 2286, "cache_hits": 7939, "stall_s": 0.7414712503086776, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1000, "installed": 1000, "late": 10, "dup": 7280, "predictor": "transition3", "m": 8}, "ssd_bytes": 11628969984, "ssd_read_s": 2.018930197576992}170{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:58:09", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 237.16158303432167, "prompt_tps": 76.00340327565287, "gen_tps": 41.39204240567406, "decode_tps_wall": 40.74555418429973, "itl_ms_p50": 24.041750002652407, "itl_ms_p95": 31.232250039465725, "itl_ms_p99": 35.08116595912725, "itl_ms_mean": 24.542554887750793, "total_s": 1.7833425409626216, "peak_mem_gb": 2.4611099, "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": 0.8191842415158976, "cache_misses": 1689, "cache_hits": 7652, "stall_s": 0.5718261183938012, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 780, "installed": 780, "late": 18, "dup": 7292, "predictor": "transition3", "m": 8}, "ssd_bytes": 8737652736, "ssd_read_s": 1.4828760590171441}171{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:58:11", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 225.4374170443043, "prompt_tps": 62.18391002055594, "gen_tps": 34.95960197813111, "decode_tps_wall": 34.41356027824513, "itl_ms_p50": 29.0010420139879, "itl_ms_p95": 36.68762499000877, "itl_ms_p99": 43.111167033202946, "itl_ms_mean": 29.058312825370756, "total_s": 2.056111125042662, "peak_mem_gb": 2.393716288, "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": 0.7700528986289539, "cache_misses": 2130, "cache_hits": 7133, "stall_s": 0.7766643203794956, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1823, "installed": 1820, "late": 53, "dup": 6225, "predictor": "transition3", "m": 8}, "ssd_bytes": 13989445632, "ssd_read_s": 2.487749135354534}172{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:58:13", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 271.2649169843644, "prompt_tps": 107.03557416160317, "gen_tps": 40.652507686234756, "decode_tps_wall": 40.01769460473437, "itl_ms_p50": 25.02000005915761, "itl_ms_p95": 31.01287502795458, "itl_ms_p99": 33.342791022732854, "itl_ms_mean": 24.988945762050292, "total_s": 1.845568499993533, "peak_mem_gb": 2.615958048, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.8019599836668028, "cache_misses": 1940, "cache_hits": 7856, "stall_s": 0.6440316166263074, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 776, "installed": 779, "late": 13, "dup": 7400, "predictor": "transition3", "m": 8}, "ssd_bytes": 9611771904, "ssd_read_s": 1.659277145168744}173{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:58:15", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 247.1101249102503, "prompt_tps": 97.25624293572443, "gen_tps": 46.4478535200105, "decode_tps_wall": 45.72248326328945, "itl_ms_p50": 21.429833956062794, "itl_ms_p95": 29.485541977919638, "itl_ms_p99": 31.328708049841225, "itl_ms_mean": 21.871078048004872, "total_s": 1.6249880419345573, "peak_mem_gb": 2.484198808, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.8527632950990616, "cache_misses": 1412, "cache_hits": 8178, "stall_s": 0.46597758354619145, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 278, "installed": 278, "late": 1, "dup": 7874, "predictor": "transition3", "m": 8}, "ssd_bytes": 5980815360, "ssd_read_s": 0.9928303242195398}174{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:58:16", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 206.4854169730097, "prompt_tps": 92.16865543735817, "gen_tps": 46.84942783945552, "decode_tps_wall": 46.11781202715002, "itl_ms_p50": 20.866874954663217, "itl_ms_p95": 27.801500051282346, "itl_ms_p99": 32.90587500669062, "itl_ms_mean": 21.683595904577817, "total_s": 1.5725519589614123, "peak_mem_gb": 2.469382252, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.8703427719821163, "cache_misses": 1218, "cache_hits": 8176, "stall_s": 0.4356100804870948, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 186, "installed": 186, "late": 1, "dup": 7910, "predictor": "transition3", "m": 8}, "ssd_bytes": 4968677376, "ssd_read_s": 0.8366910853656009}175{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:58:18", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 259.0983749832958, "prompt_tps": 112.0695527214353, "gen_tps": 40.55835547103742, "decode_tps_wall": 39.924911584296126, "itl_ms_p50": 25.12725000269711, "itl_ms_p95": 31.321458984166384, "itl_ms_p99": 32.69941604230553, "itl_ms_mean": 25.047018523475835, "total_s": 1.8370605419622734, "peak_mem_gb": 2.57750576, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.8061453654552879, "cache_misses": 1899, "cache_hits": 7897, "stall_s": 0.635601923102513, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 750, "installed": 750, "late": 14, "dup": 7426, "predictor": "transition3", "m": 8}, "ssd_bytes": 9374662656, "ssd_read_s": 1.6130917553091422}176{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:58:20", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 208.04695901460946, "prompt_tps": 91.47820347533639, "gen_tps": 47.61138101982919, "decode_tps_wall": 46.86780188914181, "itl_ms_p50": 20.600584102794528, "itl_ms_p95": 27.790957945398986, "itl_ms_p99": 30.509707983583212, "itl_ms_mean": 21.33660977669356, "total_s": 1.5522533749463037, "peak_mem_gb": 2.472921196, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.8687460080902704, "cache_misses": 1233, "cache_hits": 8161, "stall_s": 0.4324378327000886, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 193, "installed": 193, "late": 1, "dup": 7903, "predictor": "transition3", "m": 8}, "ssd_bytes": 5046534144, "ssd_read_s": 0.8399806651286781}177{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:58:21", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 235.4157919762656, "prompt_tps": 76.56716786281245, "gen_tps": 42.18419781437584, "decode_tps_wall": 41.52534796820358, "itl_ms_p50": 23.598375031724572, "itl_ms_p95": 31.155375065281987, "itl_ms_p99": 33.42212503775954, "itl_ms_mean": 24.08167658861549, "total_s": 1.7525614170590416, "peak_mem_gb": 2.453983116, "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": 0.8219676694144096, "cache_misses": 1663, "cache_hits": 7678, "stall_s": 0.5653333888622001, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 781, "installed": 781, "late": 24, "dup": 7291, "predictor": "transition3", "m": 8}, "ssd_bytes": 8649179136, "ssd_read_s": 1.4717162100132555}178{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:58:23", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 248.1124580372125, "prompt_tps": 96.84620965944319, "gen_tps": 46.84358843337115, "decode_tps_wall": 46.1120736147709, "itl_ms_p50": 21.135417046025395, "itl_ms_p95": 28.927042032592, "itl_ms_p99": 30.076291994191706, "itl_ms_mean": 21.68629431749679, "total_s": 1.6143490000395104, "peak_mem_gb": 2.491276696, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.8495307612095934, "cache_misses": 1443, "cache_hits": 8147, "stall_s": 0.4654328323667869, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 284, "installed": 284, "late": 1, "dup": 7868, "predictor": "transition3", "m": 8}, "ssd_bytes": 6111756288, "ssd_read_s": 1.0083087478997186}179{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:58:25", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 219.43383407779038, "prompt_tps": 63.89447060001168, "gen_tps": 35.38798280098549, "decode_tps_wall": 34.83523820612727, "itl_ms_p50": 28.87949999421835, "itl_ms_p95": 36.70649998821318, "itl_ms_p99": 41.11674998421222, "itl_ms_mean": 28.706564142974834, "total_s": 2.027947375085205, "peak_mem_gb": 2.363045368, "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": 0.7730756774263198, "cache_misses": 2102, "cache_hits": 7161, "stall_s": 0.7645913262385875, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1827, "installed": 1824, "late": 51, "dup": 6221, "predictor": "transition3", "m": 8}, "ssd_bytes": 13904510976, "ssd_read_s": 2.467960340087302}180{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition3", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition3 --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:58:27", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 325.7768329931423, "prompt_tps": 98.32125429356284, "gen_tps": 37.235080040657586, "decode_tps_wall": 36.65352092509186, "itl_ms_p50": 27.147167013026774, "itl_ms_p95": 33.30558305606246, "itl_ms_p99": 39.70529208891094, "itl_ms_mean": 27.282508603844143, "total_s": 2.044574875035323, "peak_mem_gb": 2.79700422, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.7754523227383863, "cache_misses": 2296, "cache_hits": 7929, "stall_s": 0.7595669905422255, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1027, "installed": 1030, "late": 16, "dup": 7253, "predictor": "transition3", "m": 8}, "ssd_bytes": 11759910912, "ssd_read_s": 2.0804126370931044}181{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:58:33", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 500.6392920622602, "prompt_tps": 64.60762933221953, "gen_tps": 29.399985349876175, "decode_tps_wall": 28.941701880083457, "itl_ms_p50": 33.084416994825006, "itl_ms_p95": 44.03874999843538, "itl_ms_p99": 54.268999956548214, "itl_ms_mean": 34.55221825390167, "total_s": 2.6774290420580655, "peak_mem_gb": 2.805456452, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.7409290953545232, "cache_misses": 2649, "cache_hits": 7576, "stall_s": 1.0457963303197175, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 6833, "installed": 6692, "late": 136, "dup": 9807, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 33343930368, "ssd_read_s": 5.073906094650738}182{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:58:36", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 261.2464999547228, "prompt_tps": 92.07228825285056, "gen_tps": 35.33606704753488, "decode_tps_wall": 34.78411063980914, "itl_ms_p50": 27.58912497665733, "itl_ms_p95": 38.29137491993606, "itl_ms_p99": 45.17758300062269, "itl_ms_mean": 28.748758602887392, "total_s": 2.0724182919366285, "peak_mem_gb": 2.926215792, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.8583941605839416, "cache_misses": 1358, "cache_hits": 8232, "stall_s": 0.569021784584038, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 5858, "installed": 5904, "late": 97, "dup": 10782, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 25434390528, "ssd_read_s": 3.4979454677086323}183{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:58:38", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 237.22312494646758, "prompt_tps": 59.15862697198368, "gen_tps": 28.355220546618618, "decode_tps_wall": 27.912391798049345, "itl_ms_p50": 35.38383392151445, "itl_ms_p95": 44.570500031113625, "itl_ms_p99": 56.429959018714726, "itl_ms_mean": 35.826381602664554, "total_s": 2.4942851659143344, "peak_mem_gb": 2.616736684, "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": 0.7543992227140235, "cache_misses": 2275, "cache_hits": 6988, "stall_s": 0.913522737333551, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 7008, "installed": 6971, "late": 175, "dup": 9632, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 32742309888, "ssd_read_s": 4.747185734449886}184{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:58:41", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 266.4266660576686, "prompt_tps": 67.70304582379634, "gen_tps": 28.715373458201757, "decode_tps_wall": 28.266818869885107, "itl_ms_p50": 34.86925002653152, "itl_ms_p95": 44.409708003513515, "itl_ms_p99": 46.300375019200146, "itl_ms_mean": 35.37716800051312, "total_s": 2.4951882500899956, "peak_mem_gb": 2.7801883, "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": 0.7618028048388823, "cache_misses": 2225, "cache_hits": 7116, "stall_s": 0.9020266789011657, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 6843, "installed": 6849, "late": 214, "dup": 9797, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 32069910528, "ssd_read_s": 4.624020031420514}185{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:58:43", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 255.62995800282806, "prompt_tps": 113.72955992030182, "gen_tps": 30.8851192655119, "decode_tps_wall": 30.402738563273047, "itl_ms_p50": 33.51187508087605, "itl_ms_p95": 43.52287494111806, "itl_ms_p99": 47.02941689174622, "itl_ms_mean": 32.89177380908754, "total_s": 2.327811707975343, "peak_mem_gb": 2.985711408, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.7978766843609637, "cache_misses": 1980, "cache_hits": 7816, "stall_s": 0.8053798165638, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 6488, "installed": 6471, "late": 136, "dup": 10152, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 29932388352, "ssd_read_s": 4.2504302660236135}186{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:58:45", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 230.87162501178682, "prompt_tps": 82.47392821817878, "gen_tps": 34.96123161300227, "decode_tps_wall": 34.41512295969738, "itl_ms_p50": 27.76787499897182, "itl_ms_p95": 41.148124961182475, "itl_ms_p99": 44.63662498164922, "itl_ms_mean": 29.056993379656756, "total_s": 2.0614622079301625, "peak_mem_gb": 2.785869828, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.8629976580796253, "cache_misses": 1287, "cache_hits": 8107, "stall_s": 0.5752399601042271, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 5915, "installed": 5938, "late": 104, "dup": 10725, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 25544097792, "ssd_read_s": 3.5176863505039364}187{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:58:47", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 329.61108395829797, "prompt_tps": 97.24129752398822, "gen_tps": 29.483137569646768, "decode_tps_wall": 29.02276827374092, "itl_ms_p50": 33.503124956041574, "itl_ms_p95": 44.09025004133582, "itl_ms_p99": 47.23362496588379, "itl_ms_mean": 34.455707001071126, "total_s": 2.500320625025779, "peak_mem_gb": 3.057408768, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.776919315403423, "cache_misses": 2281, "cache_hits": 7944, "stall_s": 0.9247209813911468, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 6707, "installed": 6689, "late": 162, "dup": 9933, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 31984975872, "ssd_read_s": 4.602123213931918}188{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:58:50", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 263.6612090282142, "prompt_tps": 68.40654715527157, "gen_tps": 28.099618415991355, "decode_tps_wall": 27.66067320788268, "itl_ms_p50": 35.72041599545628, "itl_ms_p95": 45.54595809895545, "itl_ms_p99": 56.83133297134191, "itl_ms_mean": 36.15241004745402, "total_s": 2.5412630420178175, "peak_mem_gb": 2.893237372, "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": 0.7609463654854941, "cache_misses": 2233, "cache_hits": 7108, "stall_s": 0.9181513079674914, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 6900, "installed": 6913, "late": 217, "dup": 9740, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 32112377856, "ssd_read_s": 4.674717596964911}189{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:58:53", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 243.5570830712095, "prompt_tps": 57.60687653338281, "gen_tps": 21.666581909471056, "decode_tps_wall": 21.32817394693354, "itl_ms_p50": 41.80737503338605, "itl_ms_p95": 92.98520907759666, "itl_ms_p99": 100.21262499503791, "itl_ms_mean": 46.88633928474571, "total_s": 3.1973964580101892, "peak_mem_gb": 2.64143886, "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": 0.7498650545179747, "cache_misses": 2317, "cache_hits": 6946, "stall_s": 0.9944427234586328, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 7330, "installed": 7331, "late": 68, "dup": 9310, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 34122498048, "ssd_read_s": 5.207796155358665}190{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:58:56", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 391.23033289797604, "prompt_tps": 74.2604917942824, "gen_tps": 25.224987267726544, "decode_tps_wall": 24.830876603496257, "itl_ms_p50": 40.87741696275771, "itl_ms_p95": 59.650499955751, "itl_ms_p99": 69.46666701696813, "itl_ms_mean": 40.272440476756955, "total_s": 2.9283940829336643, "peak_mem_gb": 2.985253408, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.7932829726418946, "cache_misses": 2025, "cache_hits": 7771, "stall_s": 0.9602265820140019, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 6554, "installed": 6535, "late": 103, "dup": 10086, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 30293360640, "ssd_read_s": 4.962919001583941}191{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:58:58", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 279.0375000331551, "prompt_tps": 86.19413802457292, "gen_tps": 32.43897022737971, "decode_tps_wall": 31.932382419124952, "itl_ms_p50": 29.314124956727028, "itl_ms_p95": 42.27933392394334, "itl_ms_p99": 49.144042073749006, "itl_ms_mean": 31.316172619838092, "total_s": 2.2519563750829548, "peak_mem_gb": 2.85726476, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.8607924921793535, "cache_misses": 1335, "cache_hits": 8255, "stall_s": 0.6195322249550372, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 5867, "installed": 5917, "late": 119, "dup": 10773, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 25671499776, "ssd_read_s": 3.7687663243850693}192{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:59:01", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 238.3127500070259, "prompt_tps": 79.94098405491975, "gen_tps": 32.777170610735205, "decode_tps_wall": 32.26559880647992, "itl_ms_p50": 29.837792040780187, "itl_ms_p95": 45.05579103715718, "itl_ms_p99": 47.20070806797594, "itl_ms_mean": 30.992761237679847, "total_s": 2.1908567079808563, "peak_mem_gb": 2.717012376, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.85863317010858, "cache_misses": 1328, "cache_hits": 8066, "stall_s": 0.5923730889335275, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 5842, "installed": 5818, "late": 100, "dup": 10798, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 25303449600, "ssd_read_s": 3.6563023164635524}193{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:59:03", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 284.04720895923674, "prompt_tps": 102.30366980706786, "gen_tps": 28.65849259889012, "decode_tps_wall": 28.210895248779455, "itl_ms_p50": 36.56758298166096, "itl_ms_p95": 45.93091597780585, "itl_ms_p99": 50.75000005308539, "itl_ms_mean": 35.44729761964094, "total_s": 2.5172269589966163, "peak_mem_gb": 2.941917728, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.7983870967741935, "cache_misses": 1975, "cache_hits": 7821, "stall_s": 0.8224571474129334, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 6510, "installed": 6484, "late": 119, "dup": 10130, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 29907615744, "ssd_read_s": 4.474754717899486}194{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:59:05", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 245.52683299407363, "prompt_tps": 77.56659956353012, "gen_tps": 32.47101369578315, "decode_tps_wall": 31.96403250418268, "itl_ms_p50": 29.763375059701502, "itl_ms_p95": 42.682374943979084, "itl_ms_p99": 47.419541981071234, "itl_ms_mean": 31.28516403145142, "total_s": 2.216492166975513, "peak_mem_gb": 2.789474308, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.8635299127102406, "cache_misses": 1282, "cache_hits": 8112, "stall_s": 0.5976252818945795, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 5908, "installed": 5933, "late": 109, "dup": 10732, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 25526403072, "ssd_read_s": 3.7437403085641563}195{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Describe the main differences between a Mac Studio and a Mac", "ts": "2026-11-01T01:59:08", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 261.61037501879036, "prompt_tps": 68.9416892824878, "gen_tps": 27.464052531956753, "decode_tps_wall": 27.0350746300655, "itl_ms_p50": 36.600374965928495, "itl_ms_p95": 46.41845799051225, "itl_ms_p99": 47.61833290103823, "itl_ms_mean": 36.98898611095039, "total_s": 2.591916500008665, "peak_mem_gb": 2.727021948, "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": 0.7674767155550798, "cache_misses": 2172, "cache_hits": 7169, "stall_s": 0.9160170473624021, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 6864, "installed": 6853, "late": 224, "dup": 9776, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 31938969600, "ssd_read_s": 4.686025946168229}196{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:59:10", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 261.63404202088714, "prompt_tps": 91.9208585718617, "gen_tps": 33.073144288813886, "decode_tps_wall": 32.556587418479594, "itl_ms_p50": 29.125166009180248, "itl_ms_p95": 38.984958082437515, "itl_ms_p99": 46.841374947689474, "itl_ms_mean": 30.715750000025665, "total_s": 2.196726292022504, "peak_mem_gb": 2.8326179, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.8581856100104275, "cache_misses": 1360, "cache_hits": 8230, "stall_s": 0.5971727652940899, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 5856, "installed": 5895, "late": 114, "dup": 10784, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 25653805056, "ssd_read_s": 3.7076285743387416}197{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:59:13", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 246.18675000965595, "prompt_tps": 56.988426417342126, "gen_tps": 26.874429803645274, "decode_tps_wall": 26.45488250478592, "itl_ms_p50": 37.13241696823388, "itl_ms_p95": 47.65595798380673, "itl_ms_p99": 52.26895899977535, "itl_ms_mean": 37.800205682980874, "total_s": 2.627599708037451, "peak_mem_gb": 2.603629484, "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": 0.7534276152434416, "cache_misses": 2284, "cache_hits": 6979, "stall_s": 0.9120199183234945, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 7093, "installed": 7056, "late": 134, "dup": 9547, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 33067892736, "ssd_read_s": 4.923615432111546}198{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "prev_rank", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor prev_rank --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "5edaf3e"}, "seed": 0, "warmup": false, "repeat": 1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:59:15", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 322.21449993085116, "prompt_tps": 99.48828294265009, "gen_tps": 27.120569903421792, "decode_tps_wall": 26.697235232092478, "itl_ms_p50": 35.92433303128928, "itl_ms_p95": 48.65050001535565, "itl_ms_p99": 57.47608304955065, "itl_ms_mean": 37.457062175408716, "total_s": 2.6820094169816002, "peak_mem_gb": 3.033357056, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.7722249388753056, "cache_misses": 2329, "cache_hits": 7896, "stall_s": 0.9441881097154692, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 6843, "installed": 6829, "late": 129, "dup": 9797, "predictor": "prev_rank", "m": 16}, "ssd_bytes": 32604291072, "ssd_read_s": 4.862698620767333}199{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:59:23", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 535.2712909225374, "prompt_tps": 60.273155196191404, "gen_tps": 35.65712803254787, "decode_tps_wall": 35.10161836662931, "itl_ms_p50": 27.24037505686283, "itl_ms_p95": 35.15112504828721, "itl_ms_p99": 49.79716590605676, "itl_ms_mean": 28.48871495197749, "total_s": 2.3300603328971192, "peak_mem_gb": 2.857885252, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.807921760391198, "cache_misses": 1964, "cache_hits": 8261, "stall_s": 0.8874180053826421, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 3138, "installed": 3138, "late": 206, "dup": 12838, "predictor": "transition2", "m": 16}, "ssd_bytes": 18055692288, "ssd_read_s": 3.668634384754114}200{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:59:25", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 323.38604202959687, "prompt_tps": 74.31246454899245, "gen_tps": 41.313832577221554, "decode_tps_wall": 40.66811035832868, "itl_ms_p50": 22.526916000060737, "itl_ms_p95": 36.096416995860636, "itl_ms_p99": 40.55079095996916, "itl_ms_mean": 24.589290999482195, "total_s": 1.872511374996975, "peak_mem_gb": 2.491276696, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.8741397288842544, "cache_misses": 1207, "cache_hits": 8383, "stall_s": 0.5873838003026322, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1184, "installed": 1184, "late": 59, "dup": 14568, "predictor": "transition2", "m": 16}, "ssd_bytes": 8461615104, "ssd_read_s": 1.914486086810939}201{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2026-11-01T01:59:27", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 257.2485829005018, "prompt_tps": 54.51359587139584, "gen_tps": 32.58839225386014, "decode_tps_wall": 32.079404163241, "itl_ms_p50": 30.457665910944343, "itl_ms_p95": 39.518165984191, "itl_ms_p99": 53.99866704829037, "itl_ms_mean": 31.172648809539776, "total_s": 2.2211254579015076, "peak_mem_gb": 2.52823389, "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": 0.8275936521645255, "cache_misses": 1597, "cache_hits": 7666, "stall_s": 0.7541046148398891, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 4355, "installed": 4329, "late": 334, "dup": 11205, "predictor": "transition2", "m": 16}, "ssd_bytes": 21063794688, "ssd_read_s": 3.3201080777216703}202{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2026-11-01T01:59:29", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 247.3491249838844, "prompt_tps": 72.88057963132758, "gen_tps": 38.63121945347461, "decode_tps_wall": 38.027869656437936, "itl_ms_p50": 25.281209032982588, "itl_ms_p95": 35.500874975696206, "itl_ms_p99": 40.43662501499057, "itl_ms_mean": 26.296503302301208, "total_s": 1.9040288330288604, "peak_mem_gb": 2.575077004, "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": 0.8572957927416764, "cache_misses": 1333, "cache_hits": 8008, "stall_s": 0.5783263321500272, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2955, "installed": 2959, "late": 237, "dup": 12645, "predictor": "transition2", "m": 16}, "ssd_bytes": 15174991872, "ssd_read_s": 2.4115369338542223}203{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:59:31", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 269.37645801808685, "prompt_tps": 107.81151768384234, "gen_tps": 38.007079972890885, "decode_tps_wall": 37.413449841781464, "itl_ms_p50": 26.186582981608808, "itl_ms_p95": 34.237375017255545, "itl_ms_p99": 38.784332922659814, "itl_ms_mean": 26.728355824681266, "total_s": 1.9532628749730065, "peak_mem_gb": 2.692917728, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.8428950592078399, "cache_misses": 1539, "cache_hits": 8257, "stall_s": 0.6016018842346966, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2482, "installed": 2487, "late": 127, "dup": 13318, "predictor": "transition2", "m": 16}, "ssd_bytes": 14230093824, "ssd_read_s": 2.329985384712927}204{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:59:33", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 220.89533298276365, "prompt_tps": 86.15119533674667, "gen_tps": 44.60681013316751, "decode_tps_wall": 43.910217652120465, "itl_ms_p50": 22.50249998178333, "itl_ms_p95": 29.182667029090226, "itl_ms_p99": 31.45670797675848, "itl_ms_mean": 22.773742729369257, "total_s": 1.6556411249330267, "peak_mem_gb": 2.540226668, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.8970619544390036, "cache_misses": 967, "cache_hits": 8427, "stall_s": 0.38502774434164166, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1200, "installed": 1209, "late": 41, "dup": 14448, "predictor": "transition2", "m": 16}, "ssd_bytes": 7668891648, "ssd_read_s": 1.3222372271120548}205{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:59:35", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 311.58033397514373, "prompt_tps": 102.85514074426068, "gen_tps": 34.6949340690122, "decode_tps_wall": 34.15302090529882, "itl_ms_p50": 28.45316706225276, "itl_ms_p95": 37.236333009786904, "itl_ms_p99": 49.54491602256894, "itl_ms_mean": 29.27998676230865, "total_s": 2.1562195000005886, "peak_mem_gb": 2.778785212, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.8255256723716381, "cache_misses": 1784, "cache_hits": 8441, "stall_s": 0.7007565473904833, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2929, "installed": 2937, "late": 197, "dup": 13079, "predictor": "transition2", "m": 16}, "ssd_bytes": 16679043072, "ssd_read_s": 2.7620964454254135}206{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2026-11-01T01:59:37", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 255.35870902240276, "prompt_tps": 70.5861705885171, "gen_tps": 36.12829949468941, "decode_tps_wall": 35.564010633101745, "itl_ms_p50": 27.851958060637116, "itl_ms_p95": 35.9834999544546, "itl_ms_p99": 39.681707974523306, "itl_ms_mean": 28.11831349159576, "total_s": 2.0268124589929357, "peak_mem_gb": 2.481241124, "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": 0.8545123648431645, "cache_misses": 1359, "cache_hits": 7982, "stall_s": 0.5864012060919777, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2937, "installed": 2913, "late": 233, "dup": 12663, "predictor": "transition2", "m": 16}, "ssd_bytes": 15203303424, "ssd_read_s": 2.4893748553004116}207{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2026-11-01T01:59:39", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 245.4188329866156, "prompt_tps": 57.13558902793146, "gen_tps": 33.350650641387794, "decode_tps_wall": 32.82979121022096, "itl_ms_p50": 30.520499916747212, "itl_ms_p95": 39.363207994028926, "itl_ms_p99": 51.52833310421556, "itl_ms_mean": 30.460138890212253, "total_s": 2.1644075830699876, "peak_mem_gb": 2.531807918, "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": 0.8282413904782467, "cache_misses": 1591, "cache_hits": 7672, "stall_s": 0.7381878941087052, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 4366, "installed": 4368, "late": 332, "dup": 11194, "predictor": "transition2", "m": 16}, "ssd_bytes": 21081489408, "ssd_read_s": 3.2596225037705153}208{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:59:41", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 317.87662499118596, "prompt_tps": 91.33051624775473, "gen_tps": 41.627584354211805, "decode_tps_wall": 40.97748539083085, "itl_ms_p50": 24.385959026403725, "itl_ms_p95": 30.50545894075185, "itl_ms_p99": 37.26725000888109, "itl_ms_mean": 24.403644842095673, "total_s": 1.8553062500432134, "peak_mem_gb": 2.693093168, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.8405471621069824, "cache_misses": 1562, "cache_hits": 8234, "stall_s": 0.5801819945918396, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2482, "installed": 2490, "late": 105, "dup": 13318, "predictor": "transition2", "m": 16}, "ssd_bytes": 14311489536, "ssd_read_s": 2.302620130358264}209{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:59:43", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 234.9707919638604, "prompt_tps": 102.27526981129552, "gen_tps": 45.06857507328316, "decode_tps_wall": 44.36469752953406, "itl_ms_p50": 22.080166963860393, "itl_ms_p95": 32.08154102321714, "itl_ms_p99": 35.39329208433628, "itl_ms_mean": 22.54044444536761, "total_s": 1.6550187920220196, "peak_mem_gb": 2.544360856, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.8794577685088634, "cache_misses": 1156, "cache_hits": 8434, "stall_s": 0.3957794235320762, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1237, "installed": 1251, "late": 45, "dup": 14515, "predictor": "transition2", "m": 16}, "ssd_bytes": 8468692992, "ssd_read_s": 1.3409272423014045}210{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:59:44", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 208.75837502535433, "prompt_tps": 91.15956165861617, "gen_tps": 47.000010171472624, "decode_tps_wall": 46.265963464491485, "itl_ms_p50": 21.05712506454438, "itl_ms_p95": 28.216041042469442, "itl_ms_p99": 31.784582999534905, "itl_ms_mean": 21.614161364379385, "total_s": 1.5704505409812555, "peak_mem_gb": 2.48014654, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.8946135831381733, "cache_misses": 990, "cache_hits": 8404, "stall_s": 0.3688524183817208, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1162, "installed": 1156, "late": 48, "dup": 14486, "predictor": "transition2", "m": 16}, "ssd_bytes": 7615807488, "ssd_read_s": 1.2374950614757836}211{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "reasoning", "prompt": "Q: What is the smallest positive integer divisible by 2, 3, ", "ts": "2026-11-01T01:59:46", "prompt_tokens": 29, "gen_tokens": 64, "ttft_ms": 244.17691701091826, "prompt_tps": 118.91345867928221, "gen_tps": 40.70535381982427, "decode_tps_wall": 40.069606632004536, "itl_ms_p50": 25.789916049689054, "itl_ms_p95": 31.336250016465783, "itl_ms_p99": 32.150042010471225, "itl_ms_mean": 24.956571427913058, "total_s": 1.8164409169694409, "peak_mem_gb": 2.600462816, "text": "The least common multiple (LCM) of 2, 3, 4, 5, and 6 is the smallest positive integer that is divisible by all of these numbers. The prime factorization of 2 is 2. The prime factorization of 3 is 3. T", "cache_hit_rate": 0.842486729277256, "cache_misses": 1543, "cache_hits": 8253, "stall_s": 0.5462719945935532, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2467, "installed": 2459, "late": 112, "dup": 13333, "predictor": "transition2", "m": 16}, "ssd_bytes": 14191165440, "ssd_read_s": 2.1599057024577633}212{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "code", "prompt": "Write a bash script that lists the ten largest files under a", "ts": "2026-11-01T01:59:48", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 200.30516700353473, "prompt_tps": 95.01252977045793, "gen_tps": 48.30449034823297, "decode_tps_wall": 47.55004226391123, "itl_ms_p50": 20.732083008624613, "itl_ms_p95": 26.532790972851217, "itl_ms_p99": 28.708333964459598, "itl_ms_mean": 21.03047552407675, "total_s": 1.5252251250203699, "peak_mem_gb": 2.517359468, "text": "#!/bin/bash\n\n# Specify the directory\ndir=\"/path/to/directory\"\n\n# Get the size of each file and store it in an array\nfiles=( $(ls -l \"$dir\" | awk '{print $5, $9}') )\n\n# Sort the array", "cache_hit_rate": 0.8967426016606345, "cache_misses": 970, "cache_hits": 8424, "stall_s": 0.351440568221733, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1210, "installed": 1218, "late": 40, "dup": 14438, "predictor": "transition2", "m": 16}, "ssd_bytes": 7714897920, "ssd_read_s": 1.212078574928455}213{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2026-11-01T01:59:49", "prompt_tokens": 18, "gen_tokens": 64, "ttft_ms": 230.96304200589657, "prompt_tps": 78.0410117482381, "gen_tps": 41.26642908079333, "decode_tps_wall": 40.621904175654606, "itl_ms_p50": 24.360541021451354, "itl_ms_p95": 32.594584045000374, "itl_ms_p99": 35.16958293039352, "itl_ms_mean": 24.61726057143616, "total_s": 1.7818504580063745, "peak_mem_gb": 2.482294668, "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": 0.8580451771758912, "cache_misses": 1326, "cache_hits": 8015, "stall_s": 0.5254195804009214, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2937, "installed": 2919, "late": 227, "dup": 12663, "predictor": "transition2", "m": 16}, "ssd_bytes": 15086518272, "ssd_read_s": 2.2348477555206046}214{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:59:51", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 242.7965000970289, "prompt_tps": 98.97924191302036, "gen_tps": 47.77664866967458, "decode_tps_wall": 47.030472032844315, "itl_ms_p50": 20.501999999396503, "itl_ms_p95": 29.22224998474121, "itl_ms_p99": 31.38637507800013, "itl_ms_mean": 21.262810190415216, "total_s": 1.5823535420931876, "peak_mem_gb": 2.578455916, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.8765380604796663, "cache_misses": 1184, "cache_hits": 8406, "stall_s": 0.39839902927633375, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 1222, "installed": 1246, "late": 48, "dup": 14530, "predictor": "transition2", "m": 16}, "ssd_bytes": 8514699264, "ssd_read_s": 1.3423358012223616}215{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "2026-11-01T01:59:53", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 221.3424580404535, "prompt_tps": 63.33984220635816, "gen_tps": 33.33601584236507, "decode_tps_wall": 32.81538558375421, "itl_ms_p50": 30.22895904723555, "itl_ms_p95": 39.32141698896885, "itl_ms_p99": 44.8081661015749, "itl_ms_mean": 30.47351058690794, "total_s": 2.1411736250156537, "peak_mem_gb": 2.521191086, "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": 0.8275936521645255, "cache_misses": 1597, "cache_hits": 7666, "stall_s": 0.7109807091765106, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 4383, "installed": 4352, "late": 324, "dup": 11177, "predictor": "transition2", "m": 16}, "ssd_bytes": 21162885120, "ssd_read_s": 3.1957702171057463}216{"backend": "titan", "capacity": 32, "policy": "lru", "predictor": "transition2", "m": 16, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "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 32 --predictor transition2 --m 16 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.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": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:59:55", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 314.20337490271777, "prompt_tps": 101.95785628273087, "gen_tps": 38.10694992833497, "decode_tps_wall": 37.511737309435766, "itl_ms_p50": 25.707249995321035, "itl_ms_p95": 33.5518749197945, "itl_ms_p99": 45.18720810301602, "itl_ms_mean": 26.65832274711676, "total_s": 1.9936777079710737, "peak_mem_gb": 2.842027452, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.8253300733496333, "cache_misses": 1786, "cache_hits": 8439, "stall_s": 0.6469127177260816, "hits_device": 0, "layers_complete": 0, "prefetch": {"issued": 2965, "installed": 2996, "late": 180, "dup": 13043, "predictor": "transition2", "m": 16}, "ssd_bytes": 16813522944, "ssd_read_s": 2.59259042784106}217