{ "_": "Every point the site plots, as a wtf.petrichor.drinkme.measurement record. Generated by build_points.py from the records in measurements/; `site.at` holds an RFC-6901 pointer for every number so check_points.py can re-derive it. Nothing here is interpolated, projected, or typed from memory. Only records comparable with `version` are bundled (docs/versioning.md); `excluded` names every record the build left out and why.", "lexicon": "lexicons/wtf.petrichor.drinkme.measurement.json", "version": "1.0.0", "excluded": [], "records": [ { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T08:42:26.084706+00:00", "version": "1.0.0", "environment": { "deviceClass": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "memoryBytes": 133680857088, "memoryKind": "unified", "cpuInfo": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "gpuInfo": "amdgpu 0x1586 vram=0.54GB gtt=133.14GB", "os": "Ubuntu 26.04.1 LTS", "engine": "drinkme 1.0.0 / torch 2.12.0a0+rocm7.13.0a20260411 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv twin / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "3.63", "unit": "tok/s", "samples": [ "3.63", "3.63", "3.63" ] }, { "name": "stock_prefill_tok_s", "value": "399.52", "unit": "tok/s", "samples": [ "399.11", "399.52", "399.85" ] }, { "name": "stock_ttft_s", "value": "1.3124", "unit": "s", "samples": [ "1.3124", "1.3139", "1.3117" ] }, { "name": "stock_weights_gb", "value": "62.57", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "61.392", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "3.65", "unit": "tok/s", "samples": [ "3.65", "3.65", "3.65" ] }, { "name": "twin_weights_gb", "value": "62.593", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "61.392", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "4.04", "unit": "tok/s", "samples": [ "4.04", "4.04", "4.04" ] }, { "name": "compressed_prefill_tok_s", "value": "290.09", "unit": "tok/s", "samples": [ "290.01", "290.09", "290.39" ] }, { "name": "compressed_ttft_s", "value": "1.7955", "unit": "s", "samples": [ "1.7929", "1.7996", "1.7955" ] }, { "name": "read_gb_s", "value": "240.045", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "207.617", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "45.091", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "44.07", "unit": "GB" } ], "model": { "name": "gemma-4-31B-it", "hfRepo": "google/gemma-4-31B-it", "revision": "842da3794eaa0b77d5f08bae87a17459d91ff475" }, "compression": { "profile": "sip", "bitsPerWeight": "11.269", "meanTensorBitsPerWeight": "11.283", "packedBytes": 42029991600, "residentBytes": 44900997020 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/gemma-4-31b-it_amd-ryzen-ai-max-395-w-radeon-8060s_2026-10-01.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 133.680857088, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T08:49:35.268548+00:00", "version": "1.0.0", "environment": { "deviceClass": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "memoryBytes": 133680857088, "memoryKind": "unified", "cpuInfo": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "gpuInfo": "amdgpu 0x1586 vram=0.54GB gtt=133.14GB", "os": "Ubuntu 26.04.1 LTS", "engine": "drinkme 1.0.0 / torch 2.12.0a0+rocm7.13.0a20260411 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv twin / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "3.63", "unit": "tok/s", "samples": [ "3.63", "3.63", "3.63" ] }, { "name": "stock_prefill_tok_s", "value": "397.16", "unit": "tok/s", "samples": [ "398.06", "397.16", "396.61" ] }, { "name": "stock_ttft_s", "value": "1.3148", "unit": "s", "samples": [ "1.3148", "1.3189", "1.3146" ] }, { "name": "stock_weights_gb", "value": "62.57", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "61.392", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "3.65", "unit": "tok/s", "samples": [ "3.65", "3.64", "3.65" ] }, { "name": "twin_weights_gb", "value": "62.593", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "61.392", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "2.92", "unit": "tok/s", "samples": [ "2.92", "2.92", "2.9" ] }, { "name": "compressed_prefill_tok_s", "value": "289.15", "unit": "tok/s", "samples": [ "288.87", "289.37", "289.15" ] }, { "name": "compressed_ttft_s", "value": "1.8005", "unit": "s", "samples": [ "1.804", "1.8005", "1.796" ] }, { "name": "read_gb_s", "value": "240.057", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "207.514", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "43.295", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "42.41", "unit": "GB" } ], "model": { "name": "gemma-4-31B-it", "hfRepo": "google/gemma-4-31B-it", "revision": "842da3794eaa0b77d5f08bae87a17459d91ff475" }, "compression": { "profile": "gulp", "bitsPerWeight": "10.798", "meanTensorBitsPerWeight": "10.82", "packedBytes": 40274803380, "residentBytes": 43210513664 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/gemma-4-31b-it_amd-ryzen-ai-max-395-w-radeon-8060s_2026-10-01_gulp.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 133.680857088, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-09-30T19:29:39.411687+00:00", "version": "1.0.0", "environment": { "deviceClass": "NVIDIA A100-SXM4-80GB", "memoryBytes": 85899345920, "memoryKind": "vram", "cpuInfo": "unknown", "gpuInfo": "NVIDIA A100-SXM4-80GB, 81920", "os": "Debian GNU/Linux 12 (bookworm)", "engine": "drinkme 1.0.0 / torch 2.13.0+cu130 / deltanet none / conv none / narrow-gemv triton / stock-gemv linear / raw-gemv stock / decode eager", "platform": "cuda" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "10.8", "unit": "tok/s", "samples": [ "10.89", "10.79", "10.8" ] }, { "name": "stock_prefill_tok_s", "value": "2538.79", "unit": "tok/s", "samples": [ "2538.79", "2526.05", "2551.54" ] }, { "name": "stock_ttft_s", "value": "0.2103", "unit": "s", "samples": [ "0.2148", "0.2103", "0.2063" ] }, { "name": "stock_weights_gb", "value": "62.57", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "61.392", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "8.32", "unit": "tok/s", "samples": [ "8.25", "8.32", "8.36" ] }, { "name": "twin_weights_gb", "value": "62.567", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "61.392", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "7.47", "unit": "tok/s", "samples": [ "7.47", "7.47", "7.47" ] }, { "name": "compressed_prefill_tok_s", "value": "1630.14", "unit": "tok/s", "samples": [ "1630.14", "1640.26", "1629.17" ] }, { "name": "compressed_ttft_s", "value": "0.3192", "unit": "s", "samples": [ "0.318", "0.3212", "0.3192" ] }, { "name": "read_gb_s", "value": "1731.088", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "1766.897", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "45.067", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "44.07", "unit": "GB" } ], "model": { "name": "gemma-4-31B-it", "hfRepo": "google/gemma-4-31B-it", "revision": "842da3794eaa0b77d5f08bae87a17459d91ff475" }, "compression": { "profile": "sip", "bitsPerWeight": "11.269", "meanTensorBitsPerWeight": "11.283" }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/gemma-4-31b-it_nvidia-a100-sxm4-80gb_2026-09-30.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 85.89934592, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-09-30T19:30:04.117973+00:00", "version": "1.0.0", "environment": { "deviceClass": "NVIDIA H100 80GB HBM3", "memoryBytes": 85520809984, "memoryKind": "vram", "cpuInfo": "unknown", "gpuInfo": "NVIDIA H100 80GB HBM3, 81559", "os": "Debian GNU/Linux 12 (bookworm)", "engine": "drinkme 1.0.0 / torch 2.13.0+cu130 / deltanet none / conv none / narrow-gemv triton / stock-gemv linear / raw-gemv stock / decode eager", "platform": "cuda" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "23.34", "unit": "tok/s", "samples": [ "22.95", "24.19", "23.34" ] }, { "name": "stock_prefill_tok_s", "value": "6051.12", "unit": "tok/s", "samples": [ "6029.56", "6055.34", "6051.12" ] }, { "name": "stock_ttft_s", "value": "0.087", "unit": "s", "samples": [ "0.1201", "0.087", "0.0864" ] }, { "name": "stock_weights_gb", "value": "62.57", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "61.392", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "18.32", "unit": "tok/s", "samples": [ "18.19", "18.38", "18.32" ] }, { "name": "twin_weights_gb", "value": "62.593", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "61.392", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "12.16", "unit": "tok/s", "samples": [ "12.16", "12.14", "12.16" ] }, { "name": "compressed_prefill_tok_s", "value": "3428.19", "unit": "tok/s", "samples": [ "3421.42", "3428.19", "3430.55" ] }, { "name": "compressed_ttft_s", "value": "0.1517", "unit": "s", "samples": [ "0.152", "0.1517", "0.1508" ] }, { "name": "read_gb_s", "value": "3081.913", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "3020.979", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "45.092", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "44.07", "unit": "GB" } ], "model": { "name": "gemma-4-31B-it", "hfRepo": "google/gemma-4-31B-it", "revision": "842da3794eaa0b77d5f08bae87a17459d91ff475" }, "compression": { "profile": "sip", "bitsPerWeight": "11.269", "meanTensorBitsPerWeight": "11.283" }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/gemma-4-31b-it_nvidia-h100-80gb-hbm3_2026-09-30.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 85.520809984, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T07:45:00.202081+00:00", "version": "1.0.0", "environment": { "deviceClass": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "memoryBytes": 133680857088, "memoryKind": "unified", "cpuInfo": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "gpuInfo": "amdgpu 0x1586 vram=0.54GB gtt=133.14GB", "os": "Ubuntu 26.04.1 LTS", "engine": "drinkme 1.0.0 / torch 2.12.0a0+rocm7.13.0a20260411 / deltanet fla 0.5.2 / conv fla 0.5.2 / narrow-gemv triton / stock-gemv triton / raw-gemv twin / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "13.12", "unit": "tok/s", "samples": [ "13.12", "13.12", "13.1" ] }, { "name": "stock_prefill_tok_s", "value": "1049.57", "unit": "tok/s", "samples": [ "1052.32", "1046.16", "1049.57" ] }, { "name": "stock_ttft_s", "value": "0.5204", "unit": "s", "samples": [ "0.5202", "0.5208", "0.5204" ] }, { "name": "stock_weights_gb", "value": "18.832", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "15.871", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "13.12", "unit": "tok/s", "samples": [ "13.12", "13.12", "13.11" ] }, { "name": "twin_weights_gb", "value": "18.853", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "15.871", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "15.28", "unit": "tok/s", "samples": [ "15.27", "15.29", "15.28" ] }, { "name": "compressed_prefill_tok_s", "value": "788.41", "unit": "tok/s", "samples": [ "788.41", "793.45", "784.47" ] }, { "name": "compressed_ttft_s", "value": "0.7012", "unit": "s", "samples": [ "0.7001", "0.7012", "0.7017" ] }, { "name": "read_gb_s", "value": "239.359", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "207.783", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "13.96", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "11.171", "unit": "GB" } ], "model": { "name": "MiMo-V2.6-Distill-Qwen-9B", "hfRepo": "XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B", "revision": "2367e865d009c13ac81713a2878291d33ab28177" }, "compression": { "profile": "sip", "bitsPerWeight": "11.26", "meanTensorBitsPerWeight": "11.277", "packedBytes": 11796045919, "residentBytes": 13854939892 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/mimo-v2.6-distill-qwen-9b_amd-ryzen-ai-max-395-w-radeon-8060s_2026-10-01.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 133.680857088, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T07:47:50.683728+00:00", "version": "1.0.0", "environment": { "deviceClass": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "memoryBytes": 133680857088, "memoryKind": "unified", "cpuInfo": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "gpuInfo": "amdgpu 0x1586 vram=0.54GB gtt=133.14GB", "os": "Ubuntu 26.04.1 LTS", "engine": "drinkme 1.0.0 / torch 2.12.0a0+rocm7.13.0a20260411 / deltanet fla 0.5.2 / conv fla 0.5.2 / narrow-gemv triton / stock-gemv triton / raw-gemv twin / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "13.12", "unit": "tok/s", "samples": [ "13.12", "13.11", "13.12" ] }, { "name": "stock_prefill_tok_s", "value": "1048.57", "unit": "tok/s", "samples": [ "1044.95", "1051.19", "1048.57" ] }, { "name": "stock_ttft_s", "value": "0.5206", "unit": "s", "samples": [ "0.5224", "0.5206", "0.5199" ] }, { "name": "stock_weights_gb", "value": "18.832", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "15.871", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "13.12", "unit": "tok/s", "samples": [ "13.12", "13.12", "13.12" ] }, { "name": "twin_weights_gb", "value": "18.853", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "15.871", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "14.26", "unit": "tok/s", "samples": [ "14.26", "14.25", "14.29" ] }, { "name": "compressed_prefill_tok_s", "value": "788.45", "unit": "tok/s", "samples": [ "782.96", "792.61", "788.45" ] }, { "name": "compressed_ttft_s", "value": "0.7047", "unit": "s", "samples": [ "0.7051", "0.7007", "0.7047" ] }, { "name": "read_gb_s", "value": "239.659", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "207.755", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "13.478", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "10.73", "unit": "GB" } ], "model": { "name": "MiMo-V2.6-Distill-Qwen-9B", "hfRepo": "XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B", "revision": "2367e865d009c13ac81713a2878291d33ab28177" }, "compression": { "profile": "gulp", "bitsPerWeight": "10.8", "meanTensorBitsPerWeight": "10.822", "packedBytes": 11314241021, "residentBytes": 13390173500 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/mimo-v2.6-distill-qwen-9b_amd-ryzen-ai-max-395-w-radeon-8060s_2026-10-01_gulp.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 133.680857088, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-09-30T01:28:00.530162+00:00", "version": "1.0.0", "environment": { "deviceClass": "NVIDIA L40S", "memoryBytes": 48305799168, "memoryKind": "vram", "cpuInfo": "unknown", "gpuInfo": "NVIDIA L40S, 46068", "os": "Debian GNU/Linux 12 (bookworm)", "engine": "drinkme 1.0.0 / torch 2.13.0+cu130 / deltanet fla 0.5.2 / conv torch / narrow-gemv triton / stock-gemv linear / raw-gemv stock / decode cudagraph", "platform": "cuda" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "39.57", "unit": "tok/s", "samples": [ "39.55", "39.57", "39.59" ] }, { "name": "stock_prefill_tok_s", "value": "7676.11", "unit": "tok/s", "samples": [ "7676.11", "7703.7", "7612.81" ] }, { "name": "stock_ttft_s", "value": "0.0671", "unit": "s", "samples": [ "0.0832", "0.0671", "0.0665" ] }, { "name": "stock_weights_gb", "value": "18.832", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "15.871", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "40.37", "unit": "tok/s", "samples": [ "40.38", "40.37", "40.28" ] }, { "name": "twin_weights_gb", "value": "18.828", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "15.871", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "50.18", "unit": "tok/s", "samples": [ "50.18", "50.2", "50.02" ] }, { "name": "compressed_prefill_tok_s", "value": "6003.74", "unit": "tok/s", "samples": [ "6000.33", "6037.75", "6003.74" ] }, { "name": "compressed_ttft_s", "value": "0.0945", "unit": "s", "samples": [ "0.1125", "0.0945", "0.0931" ] }, { "name": "read_gb_s", "value": "741.169", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "652.176", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "13.939", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "11.171", "unit": "GB" } ], "model": { "name": "MiMo-V2.6-Distill-Qwen-9B", "hfRepo": "XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B", "revision": "2367e865d009c13ac81713a2878291d33ab28177" }, "compression": { "profile": "sip", "bitsPerWeight": "11.26", "meanTensorBitsPerWeight": "11.277" }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/mimo-v2.6-distill-qwen-9b_nvidia-l40s_2026-09-30.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 48.305799168, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T09:59:01.945961+00:00", "version": "1.0.0", "environment": { "deviceClass": "NVIDIA L40S", "memoryBytes": 48305799168, "memoryKind": "vram", "cpuInfo": "unknown", "gpuInfo": "NVIDIA L40S, 46068", "os": "Debian GNU/Linux 12 (bookworm)", "engine": "drinkme 1.0.0 / torch 2.13.0+cu130 / deltanet fla 0.5.2 / conv torch / narrow-gemv triton / stock-gemv linear / raw-gemv stock / decode cudagraph", "platform": "cuda" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "39.64", "unit": "tok/s", "samples": [ "39.65", "39.64", "39.63" ] }, { "name": "stock_prefill_tok_s", "value": "8123.12", "unit": "tok/s", "samples": [ "8123.12", "8149.63", "8093.27" ] }, { "name": "stock_ttft_s", "value": "0.0647", "unit": "s", "samples": [ "0.0776", "0.0647", "0.0647" ] }, { "name": "stock_weights_gb", "value": "18.832", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "15.871", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "40.01", "unit": "tok/s", "samples": [ "40.01", "39.8", "40.1" ] }, { "name": "twin_weights_gb", "value": "18.828", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "15.871", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "49.3", "unit": "tok/s", "samples": [ "49.38", "49.3", "49.25" ] }, { "name": "compressed_prefill_tok_s", "value": "6162.5", "unit": "tok/s", "samples": [ "6162.5", "6158.15", "6169.39" ] }, { "name": "compressed_ttft_s", "value": "0.0906", "unit": "s", "samples": [ "0.0946", "0.0899", "0.0906" ] }, { "name": "read_gb_s", "value": "741.449", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "650.262", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "13.454", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "10.73", "unit": "GB" } ], "model": { "name": "MiMo-V2.6-Distill-Qwen-9B", "hfRepo": "XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B", "revision": "2367e865d009c13ac81713a2878291d33ab28177" }, "compression": { "profile": "gulp", "bitsPerWeight": "10.8", "meanTensorBitsPerWeight": "10.822", "packedBytes": 11314241021, "residentBytes": 13390173500 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/mimo-v2.6-distill-qwen-9b_nvidia-l40s_2026-10-01_gulp.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 48.305799168, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T08:28:45.045860+00:00", "version": "1.0.0", "environment": { "deviceClass": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "memoryBytes": 133680857088, "memoryKind": "unified", "cpuInfo": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "gpuInfo": "amdgpu 0x1586 vram=0.54GB gtt=133.14GB", "os": "Ubuntu 26.04.1 LTS", "engine": "drinkme 1.0.0 / torch 2.12.0a0+rocm7.13.0a20260411 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv twin / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "4.24", "unit": "tok/s", "samples": [ "4.24", "4.24", "4.24" ] }, { "name": "stock_prefill_tok_s", "value": "592.92", "unit": "tok/s", "samples": [ "592.92", "595.11", "591.7" ] }, { "name": "stock_ttft_s", "value": "0.8932", "unit": "s", "samples": [ "0.8932", "0.8944", "0.8913" ] }, { "name": "stock_weights_gb", "value": "59.646", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "53.017", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "4.29", "unit": "tok/s", "samples": [ "4.29", "4.29", "4.29" ] }, { "name": "twin_weights_gb", "value": "59.68", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "53.017", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "4.48", "unit": "tok/s", "samples": [ "4.48", "4.47", "4.48" ] }, { "name": "compressed_prefill_tok_s", "value": "400.92", "unit": "tok/s", "samples": [ "400.58", "400.92", "401.08" ] }, { "name": "compressed_ttft_s", "value": "1.3249", "unit": "s", "samples": [ "1.3245", "1.3285", "1.3249" ] }, { "name": "read_gb_s", "value": "233.866", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "206.753", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "42.958", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "37.48", "unit": "GB" } ], "model": { "name": "Muse-Glimmer-30B", "hfRepo": "meta-models/Muse-Glimmer-30B", "revision": "a4e59da52a7bc87ae7251dd5545c0dd437c44b68" }, "compression": { "profile": "sip", "bitsPerWeight": "11.284", "meanTensorBitsPerWeight": "11.302", "packedBytes": 39846715453, "residentBytes": 42899585080 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/muse-glimmer-30b_amd-ryzen-ai-max-395-w-radeon-8060s_2026-10-01.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 133.680857088, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T08:35:27.358022+00:00", "version": "1.0.0", "environment": { "deviceClass": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "memoryBytes": 133680857088, "memoryKind": "unified", "cpuInfo": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "gpuInfo": "amdgpu 0x1586 vram=0.54GB gtt=133.14GB", "os": "Ubuntu 26.04.1 LTS", "engine": "drinkme 1.0.0 / torch 2.12.0a0+rocm7.13.0a20260411 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv twin / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "4.29", "unit": "tok/s", "samples": [ "4.29", "4.29", "4.29" ] }, { "name": "stock_prefill_tok_s", "value": "586.7", "unit": "tok/s", "samples": [ "588.78", "585.65", "586.7" ] }, { "name": "stock_ttft_s", "value": "0.8962", "unit": "s", "samples": [ "0.9023", "0.8962", "0.8873" ] }, { "name": "stock_weights_gb", "value": "59.646", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "53.017", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "4.29", "unit": "tok/s", "samples": [ "4.29", "4.29", "4.29" ] }, { "name": "twin_weights_gb", "value": "59.68", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "53.017", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "3.14", "unit": "tok/s", "samples": [ "3.08", "3.14", "3.14" ] }, { "name": "compressed_prefill_tok_s", "value": "399.6", "unit": "tok/s", "samples": [ "399.64", "399.25", "399.6" ] }, { "name": "compressed_ttft_s", "value": "1.3208", "unit": "s", "samples": [ "1.3204", "1.3242", "1.3208" ] }, { "name": "read_gb_s", "value": "239.685", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "207.489", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "41.419", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "35.912", "unit": "GB" } ], "model": { "name": "Muse-Glimmer-30B", "hfRepo": "meta-models/Muse-Glimmer-30B", "revision": "a4e59da52a7bc87ae7251dd5545c0dd437c44b68" }, "compression": { "profile": "gulp", "bitsPerWeight": "10.797", "meanTensorBitsPerWeight": "10.838", "packedBytes": 38128257927, "residentBytes": 41240530164 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/muse-glimmer-30b_amd-ryzen-ai-max-395-w-radeon-8060s_2026-10-01_gulp.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 133.680857088, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T17:09:28.012194+00:00", "version": "1.0.0", "environment": { "deviceClass": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "memoryBytes": 133680857088, "memoryKind": "unified", "cpuInfo": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "gpuInfo": "amdgpu 0x1586 vram=0.54GB gtt=133.14GB", "os": "Ubuntu 26.04.1 LTS", "engine": "drinkme 1.0.0 / torch 2.12.0a0+rocm7.13.0a20260411 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv twin / decode eager", "platform": "rocm" }, "metrics": [ { "name": "compressed_decode_tok_s", "value": "1.91", "unit": "tok/s", "samples": [ "1.91", "1.91", "1.91" ] }, { "name": "compressed_prefill_tok_s", "value": "113.9", "unit": "tok/s", "samples": [ "113.82", "113.9", "113.94" ] }, { "name": "compressed_ttft_s", "value": "4.5733", "unit": "s", "samples": [ "4.6716", "4.5672", "4.5733" ] }, { "name": "read_gb_s", "value": "239.823", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "204.505", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "103.482", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "100.829", "unit": "GB" } ], "model": { "name": "Qwen2.5-72B", "hfRepo": "Qwen/Qwen2.5-72B-Instruct", "revision": "495f39366efef23836d0cfae4fbe635880d2be31" }, "compression": { "profile": "sip", "bitsPerWeight": "11.288", "meanTensorBitsPerWeight": "11.298", "packedBytes": 100828371245, "residentBytes": 103324581364 }, "stock": { "outcome": "skipped_predicted_nonfit", "budgetBytes": 133680857088 }, "site": { "source": "measurements/qwen2.5-72b_amd-ryzen-ai-max-395-w-radeon-8060s_2026-10-01.json", "at": { "compressed_decode_tok_s": "/metrics/0/value", "compressed_prefill_tok_s": "/metrics/1/value", "compressed_ttft_s": "/metrics/2/value", "read_gb_s": "/metrics/3/value", "copy_gb_s": "/metrics/4/value", "compressed_weights_gb": "/metrics/5/value", "compressed_decode_read_gb": "/metrics/6/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "stock.budgetBytes": "/stock/budgetBytes", "stockError": "/raw/stock_error", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 133.680857088, "harness": "drinkme bench", "stockError": "skipped by bench's stock-arm fit check (predicted OOM under the stream loader's transient)" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T17:52:10.552285+00:00", "version": "1.0.0", "environment": { "deviceClass": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "memoryBytes": 133680857088, "memoryKind": "unified", "cpuInfo": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "gpuInfo": "amdgpu 0x1586 vram=0.54GB gtt=133.14GB", "os": "Ubuntu 26.04.1 LTS", "engine": "drinkme 1.0.0 / torch 2.12.0a0+rocm7.13.0a20260411 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv twin / decode eager", "platform": "rocm" }, "metrics": [ { "name": "compressed_decode_tok_s", "value": "1.39", "unit": "tok/s", "samples": [ "1.39", "1.39", "1.4" ] }, { "name": "compressed_prefill_tok_s", "value": "106.64", "unit": "tok/s", "samples": [ "107.62", "98.87", "106.64" ] }, { "name": "compressed_ttft_s", "value": "4.861", "unit": "s", "samples": [ "4.8539", "4.861", "5.276" ] }, { "name": "read_gb_s", "value": "237.456", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "200.715", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "99.112", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "96.556", "unit": "GB" } ], "model": { "name": "Qwen2.5-72B", "hfRepo": "Qwen/Qwen2.5-72B-Instruct", "revision": "495f39366efef23836d0cfae4fbe635880d2be31" }, "compression": { "profile": "gulp", "bitsPerWeight": "10.794", "meanTensorBitsPerWeight": "10.822", "packedBytes": 96415147283, "residentBytes": 99051432776 }, "stock": { "outcome": "skipped_predicted_nonfit", "budgetBytes": 133680857088 }, "site": { "source": "measurements/qwen2.5-72b_amd-ryzen-ai-max-395-w-radeon-8060s_2026-10-01_gulp.json", "at": { "compressed_decode_tok_s": "/metrics/0/value", "compressed_prefill_tok_s": "/metrics/1/value", "compressed_ttft_s": "/metrics/2/value", "read_gb_s": "/metrics/3/value", "copy_gb_s": "/metrics/4/value", "compressed_weights_gb": "/metrics/5/value", "compressed_decode_read_gb": "/metrics/6/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "stock.budgetBytes": "/stock/budgetBytes", "stockError": "/raw/stock_error", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 133.680857088, "harness": "drinkme bench", "stockError": "skipped by bench's stock-arm fit check (predicted OOM under the stream loader's transient)" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T07:34:14.678953+00:00", "version": "1.0.0", "environment": { "deviceClass": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "memoryBytes": 133680857088, "memoryKind": "unified", "cpuInfo": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "gpuInfo": "amdgpu 0x1586 vram=0.54GB gtt=133.14GB", "os": "Ubuntu 26.04.1 LTS", "engine": "drinkme 1.0.0 / torch 2.12.0a0+rocm7.13.0a20260411 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv twin / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "100.26", "unit": "tok/s", "samples": [ "100.26", "100.15", "100.3" ] }, { "name": "stock_prefill_tok_s", "value": "11642.95", "unit": "tok/s", "samples": [ "11642.95", "11610.91", "11821.64" ] }, { "name": "stock_ttft_s", "value": "0.0469", "unit": "s", "samples": [ "0.0469", "0.0471", "0.0468" ] }, { "name": "stock_weights_gb", "value": "1.192", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "1.192", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "100.35", "unit": "tok/s", "samples": [ "100.41", "100.35", "100.07" ] }, { "name": "twin_weights_gb", "value": "1.226", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "1.192", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "100.7", "unit": "tok/s", "samples": [ "100.23", "100.7", "100.82" ] }, { "name": "compressed_prefill_tok_s", "value": "12624.01", "unit": "tok/s", "samples": [ "12610.37", "12682.31", "12624.01" ] }, { "name": "compressed_ttft_s", "value": "0.0443", "unit": "s", "samples": [ "0.045", "0.0443", "0.044" ] }, { "name": "read_gb_s", "value": "238.706", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "207.349", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "0.973", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "0.934", "unit": "GB" } ], "model": { "name": "Qwen3-0.6B", "hfRepo": "Qwen/Qwen3-0.6B", "revision": "c1899de289a04d12100db370d81485cdf75e47ca" }, "compression": { "profile": "sip", "bitsPerWeight": "11.31", "meanTensorBitsPerWeight": "11.316", "packedBytes": 622929168, "residentBytes": 1245571120 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-0.6b_amd-ryzen-ai-max-395-w-radeon-8060s_2026-10-01.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 133.680857088, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T07:35:03.528929+00:00", "version": "1.0.0", "environment": { "deviceClass": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "memoryBytes": 133680857088, "memoryKind": "unified", "cpuInfo": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "gpuInfo": "amdgpu 0x1586 vram=0.54GB gtt=133.14GB", "os": "Ubuntu 26.04.1 LTS", "engine": "drinkme 1.0.0 / torch 2.12.0a0+rocm7.13.0a20260411 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv twin / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "100.34", "unit": "tok/s", "samples": [ "100.34", "100.26", "100.35" ] }, { "name": "stock_prefill_tok_s", "value": "11839.24", "unit": "tok/s", "samples": [ "11712.55", "11845.57", "11839.24" ] }, { "name": "stock_ttft_s", "value": "0.047", "unit": "s", "samples": [ "0.047", "0.047", "0.047" ] }, { "name": "stock_weights_gb", "value": "1.192", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "1.192", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "100.27", "unit": "tok/s", "samples": [ "100.34", "100.27", "100.27" ] }, { "name": "twin_weights_gb", "value": "1.226", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "1.192", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "92.97", "unit": "tok/s", "samples": [ "93.1", "92.77", "92.97" ] }, { "name": "compressed_prefill_tok_s", "value": "12699.94", "unit": "tok/s", "samples": [ "12548.4", "12699.94", "12752.58" ] }, { "name": "compressed_ttft_s", "value": "0.0444", "unit": "s", "samples": [ "0.0447", "0.0444", "0.0443" ] }, { "name": "read_gb_s", "value": "239.736", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "207.684", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "0.948", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "0.911", "unit": "GB" } ], "model": { "name": "Qwen3-0.6B", "hfRepo": "Qwen/Qwen3-0.6B", "revision": "c1899de289a04d12100db370d81485cdf75e47ca" }, "compression": { "profile": "gulp", "bitsPerWeight": "10.874", "meanTensorBitsPerWeight": "10.886", "packedBytes": 598928188, "residentBytes": 1222550636 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-0.6b_amd-ryzen-ai-max-395-w-radeon-8060s_2026-10-01_gulp.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 133.680857088, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T07:35:48.824664+00:00", "version": "1.0.0", "environment": { "deviceClass": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "memoryBytes": 133680857088, "memoryKind": "unified", "cpuInfo": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "gpuInfo": "amdgpu 0x1586 vram=0.54GB gtt=133.14GB", "os": "Ubuntu 26.04.1 LTS", "engine": "drinkme 1.0.0 / torch 2.12.0a0+rocm7.13.0a20260411 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv twin / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "50.44", "unit": "tok/s", "samples": [ "50.45", "50.44", "50.39" ] }, { "name": "stock_prefill_tok_s", "value": "4529.56", "unit": "tok/s", "samples": [ "4517.58", "4573.11", "4529.56" ] }, { "name": "stock_ttft_s", "value": "0.1261", "unit": "s", "samples": [ "0.1261", "0.127", "0.1244" ] }, { "name": "stock_weights_gb", "value": "3.442", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "3.441", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "50.44", "unit": "tok/s", "samples": [ "50.44", "50.47", "50.44" ] }, { "name": "twin_weights_gb", "value": "3.475", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "3.441", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "53.72", "unit": "tok/s", "samples": [ "53.7", "53.73", "53.72" ] }, { "name": "compressed_prefill_tok_s", "value": "4055.66", "unit": "tok/s", "samples": [ "4015.15", "4057.13", "4055.66" ] }, { "name": "compressed_ttft_s", "value": "0.1363", "unit": "s", "samples": [ "0.1345", "0.1365", "0.1363" ] }, { "name": "read_gb_s", "value": "238.782", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "207.714", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "2.647", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "2.607", "unit": "GB" } ], "model": { "name": "Qwen3-1.7B", "hfRepo": "Qwen/Qwen3-1.7B", "revision": "70d244cc86ccca08cf5af4e1e306ecf908b1ad5e" }, "compression": { "profile": "sip", "bitsPerWeight": "11.263", "meanTensorBitsPerWeight": "11.273", "packedBytes": 1984349124, "residentBytes": 3229437676 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-1.7b_amd-ryzen-ai-max-395-w-radeon-8060s_2026-10-01.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 133.680857088, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T07:36:46.202932+00:00", "version": "1.0.0", "environment": { "deviceClass": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "memoryBytes": 133680857088, "memoryKind": "unified", "cpuInfo": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "gpuInfo": "amdgpu 0x1586 vram=0.54GB gtt=133.14GB", "os": "Ubuntu 26.04.1 LTS", "engine": "drinkme 1.0.0 / torch 2.12.0a0+rocm7.13.0a20260411 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv twin / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "50.42", "unit": "tok/s", "samples": [ "50.26", "50.42", "50.44" ] }, { "name": "stock_prefill_tok_s", "value": "4510.74", "unit": "tok/s", "samples": [ "4483.18", "4519.14", "4510.74" ] }, { "name": "stock_ttft_s", "value": "0.1268", "unit": "s", "samples": [ "0.127", "0.1268", "0.1261" ] }, { "name": "stock_weights_gb", "value": "3.442", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "3.441", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "50.5", "unit": "tok/s", "samples": [ "50.5", "50.5", "50.47" ] }, { "name": "twin_weights_gb", "value": "3.475", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "3.441", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "51.36", "unit": "tok/s", "samples": [ "51.33", "51.36", "51.44" ] }, { "name": "compressed_prefill_tok_s", "value": "3828.8", "unit": "tok/s", "samples": [ "3828.8", "3921.37", "3823.93" ] }, { "name": "compressed_ttft_s", "value": "0.1501", "unit": "s", "samples": [ "0.1506", "0.1501", "0.1501" ] }, { "name": "read_gb_s", "value": "240.03", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "207.701", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "2.571", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "2.531", "unit": "GB" } ], "model": { "name": "Qwen3-1.7B", "hfRepo": "Qwen/Qwen3-1.7B", "revision": "70d244cc86ccca08cf5af4e1e306ecf908b1ad5e" }, "compression": { "profile": "gulp", "bitsPerWeight": "10.816", "meanTensorBitsPerWeight": "10.833", "packedBytes": 1905699240, "residentBytes": 3153660684 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-1.7b_amd-ryzen-ai-max-395-w-radeon-8060s_2026-10-01_gulp.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 133.680857088, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-09-29T11:21:53.945601+00:00", "version": "1.0.0", "environment": { "deviceClass": "Apple M4", "memoryBytes": 25769803776, "memoryKind": "unified", "cpuInfo": "Apple M4", "gpuInfo": "Apple M4", "os": "macOS 15.7.7", "engine": "drinkme 1.0.0 / mlx 0.32.2", "platform": "metal" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "27.71", "unit": "tok/s", "samples": [ "27.64", "27.71", "27.81" ] }, { "name": "stock_prefill_tok_s", "value": "1212.32", "unit": "tok/s", "samples": [ "1212.32", "1213.83", "1210.23" ] }, { "name": "stock_ttft_s", "value": "0.4222", "unit": "s", "samples": [ "0.4219", "0.4222", "0.4222" ] }, { "name": "stock_weights_gb", "value": "3.441", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "33.59", "unit": "tok/s", "samples": [ "33.62", "33.58", "33.59" ] }, { "name": "compressed_prefill_tok_s", "value": "1094.28", "unit": "tok/s", "samples": [ "1094.28", "1094.82", "1094.09" ] }, { "name": "compressed_ttft_s", "value": "0.4684", "unit": "s", "samples": [ "0.4685", "0.4684", "0.4679" ] }, { "name": "read_gb_s", "value": "104.114", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "98.094", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "2.607", "unit": "GB" } ], "model": { "name": "Qwen3-1.7B", "hfRepo": "Qwen/Qwen3-1.7B", "revision": "70d244cc86ccca08cf5af4e1e306ecf908b1ad5e" }, "compression": { "profile": "sip", "bitsPerWeight": "11.263", "meanTensorBitsPerWeight": "11.273", "packedBytes": 1984349124, "residentBytes": 3229437676 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-1.7b_apple-m4_2026-09-29_2.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "compressed_decode_tok_s": "/metrics/4/value", "compressed_prefill_tok_s": "/metrics/5/value", "compressed_ttft_s": "/metrics/6/value", "read_gb_s": "/metrics/7/value", "copy_gb_s": "/metrics/8/value", "compressed_weights_gb": "/metrics/9/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 25.769803776, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T08:51:12.153068+00:00", "version": "1.0.0", "environment": { "deviceClass": "Apple M4", "memoryBytes": 25769803776, "memoryKind": "unified", "cpuInfo": "Apple M4", "gpuInfo": "Apple M4", "os": "macOS 15.7.7", "engine": "drinkme 1.0.0 / mlx 0.32.2", "platform": "metal" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "27.45", "unit": "tok/s", "samples": [ "27.64", "27.36", "27.45" ] }, { "name": "stock_prefill_tok_s", "value": "1212.96", "unit": "tok/s", "samples": [ "1213.1", "1212.96", "1211.42" ] }, { "name": "stock_ttft_s", "value": "0.4226", "unit": "s", "samples": [ "0.4218", "0.4226", "0.4226" ] }, { "name": "stock_weights_gb", "value": "3.441", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "28.73", "unit": "tok/s", "samples": [ "28.73", "28.69", "28.74" ] }, { "name": "compressed_prefill_tok_s", "value": "1094.79", "unit": "tok/s", "samples": [ "1094.61", "1094.79", "1095.91" ] }, { "name": "compressed_ttft_s", "value": "0.4674", "unit": "s", "samples": [ "0.4679", "0.4674", "0.4672" ] }, { "name": "read_gb_s", "value": "104.211", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "96.332", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "2.529", "unit": "GB" } ], "model": { "name": "Qwen3-1.7B", "hfRepo": "Qwen/Qwen3-1.7B", "revision": "70d244cc86ccca08cf5af4e1e306ecf908b1ad5e" }, "compression": { "profile": "gulp", "bitsPerWeight": "10.816", "meanTensorBitsPerWeight": "10.833", "packedBytes": 1905699240, "residentBytes": 3153660684 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-1.7b_apple-m4_2026-10-01_gulp.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "compressed_decode_tok_s": "/metrics/4/value", "compressed_prefill_tok_s": "/metrics/5/value", "compressed_ttft_s": "/metrics/6/value", "read_gb_s": "/metrics/7/value", "copy_gb_s": "/metrics/8/value", "compressed_weights_gb": "/metrics/9/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 25.769803776, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T08:57:00.004051+00:00", "version": "1.0.0", "environment": { "deviceClass": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "memoryBytes": 133680857088, "memoryKind": "unified", "cpuInfo": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "gpuInfo": "amdgpu 0x1586 vram=0.54GB gtt=133.14GB", "os": "Ubuntu 26.04.1 LTS", "engine": "drinkme 1.0.0 / torch 2.12.0a0+rocm7.13.0a20260411 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv twin / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "3.59", "unit": "tok/s", "samples": [ "3.59", "3.59", "3.59" ] }, { "name": "stock_prefill_tok_s", "value": "399.71", "unit": "tok/s", "samples": [ "399.03", "400.1", "399.71" ] }, { "name": "stock_ttft_s", "value": "1.2958", "unit": "s", "samples": [ "1.2958", "1.3008", "1.2945" ] }, { "name": "stock_weights_gb", "value": "65.525", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "63.967", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "3.59", "unit": "tok/s", "samples": [ "3.59", "3.59", "3.59" ] }, { "name": "twin_weights_gb", "value": "65.558", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "63.967", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "4.28", "unit": "tok/s", "samples": [ "4.29", "4.27", "4.28" ] }, { "name": "compressed_prefill_tok_s", "value": "285.96", "unit": "tok/s", "samples": [ "285.32", "285.96", "286.11" ] }, { "name": "compressed_ttft_s", "value": "1.8271", "unit": "s", "samples": [ "1.8271", "1.8219", "1.8279" ] }, { "name": "read_gb_s", "value": "239.543", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "206.907", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "46.974", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "45.274", "unit": "GB" } ], "model": { "name": "Qwen3-32B", "hfRepo": "Qwen/Qwen3-32B", "revision": "9216db5781bf21249d130ec9da846c4624c16137" }, "compression": { "profile": "sip", "bitsPerWeight": "11.324", "meanTensorBitsPerWeight": "11.304", "packedBytes": 45273501469, "residentBytes": 46831093764 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-32b_amd-ryzen-ai-max-395-w-radeon-8060s_2026-10-01.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 133.680857088, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T09:03:36.371043+00:00", "version": "1.0.0", "environment": { "deviceClass": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "memoryBytes": 133680857088, "memoryKind": "unified", "cpuInfo": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "gpuInfo": "amdgpu 0x1586 vram=0.54GB gtt=133.14GB", "os": "Ubuntu 26.04.1 LTS", "engine": "drinkme 1.0.0 / torch 2.12.0a0+rocm7.13.0a20260411 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv twin / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "3.59", "unit": "tok/s", "samples": [ "3.55", "3.59", "3.59" ] }, { "name": "stock_prefill_tok_s", "value": "401.94", "unit": "tok/s", "samples": [ "401.56", "401.94", "402.45" ] }, { "name": "stock_ttft_s", "value": "1.2884", "unit": "s", "samples": [ "1.2884", "1.2916", "1.2861" ] }, { "name": "stock_weights_gb", "value": "65.525", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "63.967", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "3.59", "unit": "tok/s", "samples": [ "3.59", "3.59", "3.59" ] }, { "name": "twin_weights_gb", "value": "65.558", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "63.967", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "3.96", "unit": "tok/s", "samples": [ "3.98", "3.91", "3.96" ] }, { "name": "compressed_prefill_tok_s", "value": "287.59", "unit": "tok/s", "samples": [ "287.59", "287.15", "287.78" ] }, { "name": "compressed_ttft_s", "value": "1.8", "unit": "s", "samples": [ "1.8", "1.8065", "1.7979" ] }, { "name": "read_gb_s", "value": "239.415", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "207.453", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "45.149", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "43.532", "unit": "GB" } ], "model": { "name": "Qwen3-32B", "hfRepo": "Qwen/Qwen3-32B", "revision": "9216db5781bf21249d130ec9da846c4624c16137" }, "compression": { "profile": "gulp", "bitsPerWeight": "10.873", "meanTensorBitsPerWeight": "10.852", "packedBytes": 43468435119, "residentBytes": 45088770936 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-32b_amd-ryzen-ai-max-395-w-radeon-8060s_2026-10-01_gulp.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 133.680857088, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T07:37:44.309486+00:00", "version": "1.0.0", "environment": { "deviceClass": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "memoryBytes": 133680857088, "memoryKind": "unified", "cpuInfo": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "gpuInfo": "amdgpu 0x1586 vram=0.54GB gtt=133.14GB", "os": "Ubuntu 26.04.1 LTS", "engine": "drinkme 1.0.0 / torch 2.12.0a0+rocm7.13.0a20260411 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv twin / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "24.52", "unit": "tok/s", "samples": [ "24.53", "24.52", "24.52" ] }, { "name": "stock_prefill_tok_s", "value": "3091.97", "unit": "tok/s", "samples": [ "3091.97", "3091.77", "3099.92" ] }, { "name": "stock_ttft_s", "value": "0.1715", "unit": "s", "samples": [ "0.172", "0.1715", "0.1711" ] }, { "name": "stock_weights_gb", "value": "8.102", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "8.045", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "24.53", "unit": "tok/s", "samples": [ "24.53", "24.53", "24.54" ] }, { "name": "twin_weights_gb", "value": "8.135", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "8.045", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "23.82", "unit": "tok/s", "samples": [ "23.83", "23.82", "23.82" ] }, { "name": "compressed_prefill_tok_s", "value": "2402.49", "unit": "tok/s", "samples": [ "2402.43", "2402.49", "2403.18" ] }, { "name": "compressed_ttft_s", "value": "0.2199", "unit": "s", "samples": [ "0.2206", "0.2198", "0.2199" ] }, { "name": "read_gb_s", "value": "238.722", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "207.641", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "6.02", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "5.92", "unit": "GB" } ], "model": { "name": "Qwen3-4B", "hfRepo": "Qwen/Qwen3-4B", "revision": "1cfa9a7208912126459214e8b04321603b3df60c" }, "compression": { "profile": "sip", "bitsPerWeight": "11.32", "meanTensorBitsPerWeight": "11.314", "packedBytes": 5141508116, "residentBytes": 5920045376 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-4b_amd-ryzen-ai-max-395-w-radeon-8060s_2026-10-01.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 133.680857088, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T07:39:14.519600+00:00", "version": "1.0.0", "environment": { "deviceClass": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "memoryBytes": 133680857088, "memoryKind": "unified", "cpuInfo": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "gpuInfo": "amdgpu 0x1586 vram=0.54GB gtt=133.14GB", "os": "Ubuntu 26.04.1 LTS", "engine": "drinkme 1.0.0 / torch 2.12.0a0+rocm7.13.0a20260411 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv twin / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "24.49", "unit": "tok/s", "samples": [ "24.53", "24.49", "24.49" ] }, { "name": "stock_prefill_tok_s", "value": "3104.29", "unit": "tok/s", "samples": [ "3088.1", "3104.29", "3113.98" ] }, { "name": "stock_ttft_s", "value": "0.1714", "unit": "s", "samples": [ "0.1727", "0.171", "0.1714" ] }, { "name": "stock_weights_gb", "value": "8.102", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "8.045", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "24.49", "unit": "tok/s", "samples": [ "24.49", "24.49", "24.5" ] }, { "name": "twin_weights_gb", "value": "8.135", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "8.045", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "17.8", "unit": "tok/s", "samples": [ "17.82", "17.8", "17.8" ] }, { "name": "compressed_prefill_tok_s", "value": "2352.14", "unit": "tok/s", "samples": [ "2354.51", "2345.47", "2352.14" ] }, { "name": "compressed_ttft_s", "value": "0.224", "unit": "s", "samples": [ "0.2253", "0.224", "0.2237" ] }, { "name": "read_gb_s", "value": "239.807", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "207.427", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "5.8", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "5.733", "unit": "GB" } ], "model": { "name": "Qwen3-4B", "hfRepo": "Qwen/Qwen3-4B", "revision": "1cfa9a7208912126459214e8b04321603b3df60c" }, "compression": { "profile": "gulp", "bitsPerWeight": "10.89", "meanTensorBitsPerWeight": "10.884", "packedBytes": 4946199508, "residentBytes": 5733001536 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-4b_amd-ryzen-ai-max-395-w-radeon-8060s_2026-10-01_gulp.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 133.680857088, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-09-29T11:28:30.513256+00:00", "version": "1.0.0", "environment": { "deviceClass": "Apple M4", "memoryBytes": 25769803776, "memoryKind": "unified", "cpuInfo": "Apple M4", "gpuInfo": "Apple M4", "os": "macOS 15.7.7", "engine": "drinkme 1.0.0 / mlx 0.32.2", "platform": "metal" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "12.08", "unit": "tok/s", "samples": [ "12.08", "12.05", "12.08" ] }, { "name": "stock_prefill_tok_s", "value": "471.34", "unit": "tok/s", "samples": [ "471.52", "471.34", "470.81" ] }, { "name": "stock_ttft_s", "value": "1.0875", "unit": "s", "samples": [ "1.0875", "1.0887", "1.0866" ] }, { "name": "stock_weights_gb", "value": "8.045", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "15.24", "unit": "tok/s", "samples": [ "15.22", "15.24", "15.28" ] }, { "name": "compressed_prefill_tok_s", "value": "422.44", "unit": "tok/s", "samples": [ "422.44", "421.75", "422.58" ] }, { "name": "compressed_ttft_s", "value": "1.2111", "unit": "s", "samples": [ "1.2107", "1.2142", "1.2111" ] }, { "name": "read_gb_s", "value": "103.939", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "82.572", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "5.92", "unit": "GB" } ], "model": { "name": "Qwen3-4B", "hfRepo": "Qwen/Qwen3-4B", "revision": "1cfa9a7208912126459214e8b04321603b3df60c" }, "compression": { "profile": "sip", "bitsPerWeight": "11.32", "meanTensorBitsPerWeight": "11.314", "packedBytes": 5141508116, "residentBytes": 5920045376 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-4b_apple-m4_2026-09-29_2.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "compressed_decode_tok_s": "/metrics/4/value", "compressed_prefill_tok_s": "/metrics/5/value", "compressed_ttft_s": "/metrics/6/value", "read_gb_s": "/metrics/7/value", "copy_gb_s": "/metrics/8/value", "compressed_weights_gb": "/metrics/9/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 25.769803776, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T08:45:50.715967+00:00", "version": "1.0.0", "environment": { "deviceClass": "Apple M4", "memoryBytes": 25769803776, "memoryKind": "unified", "cpuInfo": "Apple M4", "gpuInfo": "Apple M4", "os": "macOS 15.7.7", "engine": "drinkme 1.0.0 / mlx 0.32.2", "platform": "metal" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "12.0", "unit": "tok/s", "samples": [ "12.0", "12.01", "12.0" ] }, { "name": "stock_prefill_tok_s", "value": "471.73", "unit": "tok/s", "samples": [ "471.95", "471.56", "471.73" ] }, { "name": "stock_ttft_s", "value": "1.0862", "unit": "s", "samples": [ "1.0862", "1.0843", "1.088" ] }, { "name": "stock_weights_gb", "value": "8.045", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "12.3", "unit": "tok/s", "samples": [ "12.3", "12.29", "12.3" ] }, { "name": "compressed_prefill_tok_s", "value": "423.55", "unit": "tok/s", "samples": [ "422.75", "424.03", "423.55" ] }, { "name": "compressed_ttft_s", "value": "1.2211", "unit": "s", "samples": [ "1.211", "1.2211", "1.2291" ] }, { "name": "read_gb_s", "value": "102.846", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "96.973", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "5.725", "unit": "GB" } ], "model": { "name": "Qwen3-4B", "hfRepo": "Qwen/Qwen3-4B", "revision": "1cfa9a7208912126459214e8b04321603b3df60c" }, "compression": { "profile": "gulp", "bitsPerWeight": "10.89", "meanTensorBitsPerWeight": "10.884", "packedBytes": 4946199508, "residentBytes": 5733001536 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-4b_apple-m4_2026-10-01_gulp.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "compressed_decode_tok_s": "/metrics/4/value", "compressed_prefill_tok_s": "/metrics/5/value", "compressed_ttft_s": "/metrics/6/value", "read_gb_s": "/metrics/7/value", "copy_gb_s": "/metrics/8/value", "compressed_weights_gb": "/metrics/9/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 25.769803776, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-09-30T01:35:33.025730+00:00", "version": "1.0.0", "environment": { "deviceClass": "NVIDIA L40S", "memoryBytes": 48305799168, "memoryKind": "vram", "cpuInfo": "unknown", "gpuInfo": "NVIDIA L40S, 46068", "os": "Debian GNU/Linux 12 (bookworm)", "engine": "drinkme 1.0.0 / torch 2.13.0+cu130 / deltanet none / conv none / narrow-gemv triton / stock-gemv linear / raw-gemv stock / decode cudagraph", "platform": "cuda" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "67.99", "unit": "tok/s", "samples": [ "68.02", "67.97", "67.99" ] }, { "name": "stock_prefill_tok_s", "value": "12673.18", "unit": "tok/s", "samples": [ "12427.71", "12699.34", "12673.18" ] }, { "name": "stock_ttft_s", "value": "0.0456", "unit": "s", "samples": [ "0.0826", "0.0449", "0.0456" ] }, { "name": "stock_weights_gb", "value": "8.102", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "8.045", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "70.69", "unit": "tok/s", "samples": [ "70.75", "70.69", "70.69" ] }, { "name": "twin_weights_gb", "value": "8.11", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "8.045", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "73.48", "unit": "tok/s", "samples": [ "73.55", "73.48", "73.46" ] }, { "name": "compressed_prefill_tok_s", "value": "9006.47", "unit": "tok/s", "samples": [ "8815.04", "9006.47", "9015.62" ] }, { "name": "compressed_ttft_s", "value": "0.0594", "unit": "s", "samples": [ "0.081", "0.0594", "0.0586" ] }, { "name": "read_gb_s", "value": "738.368", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "653.096", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "5.995", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "5.92", "unit": "GB" } ], "model": { "name": "Qwen3-4B", "hfRepo": "Qwen/Qwen3-4B", "revision": "1cfa9a7208912126459214e8b04321603b3df60c" }, "compression": { "profile": "sip", "bitsPerWeight": "11.32", "meanTensorBitsPerWeight": "11.313" }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-4b_nvidia-l40s_2026-09-30.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 48.305799168, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T09:55:45.286772+00:00", "version": "1.0.0", "environment": { "deviceClass": "NVIDIA L40S", "memoryBytes": 48305799168, "memoryKind": "vram", "cpuInfo": "unknown", "gpuInfo": "NVIDIA L40S, 46068", "os": "Debian GNU/Linux 12 (bookworm)", "engine": "drinkme 1.0.0 / torch 2.13.0+cu130 / deltanet none / conv none / narrow-gemv triton / stock-gemv linear / raw-gemv stock / decode cudagraph", "platform": "cuda" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "69.3", "unit": "tok/s", "samples": [ "69.32", "69.28", "69.3" ] }, { "name": "stock_prefill_tok_s", "value": "16363.53", "unit": "tok/s", "samples": [ "16905.33", "16363.53", "16342.11" ] }, { "name": "stock_ttft_s", "value": "0.0328", "unit": "s", "samples": [ "0.0336", "0.0328", "0.0325" ] }, { "name": "stock_weights_gb", "value": "8.102", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "8.045", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "70.91", "unit": "tok/s", "samples": [ "70.91", "70.95", "70.9" ] }, { "name": "twin_weights_gb", "value": "8.11", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "8.045", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "40.99", "unit": "tok/s", "samples": [ "41.0", "40.99", "40.99" ] }, { "name": "compressed_prefill_tok_s", "value": "10643.86", "unit": "tok/s", "samples": [ "10643.86", "10645.92", "10508.34" ] }, { "name": "compressed_ttft_s", "value": "0.0494", "unit": "s", "samples": [ "0.053", "0.049", "0.0494" ] }, { "name": "read_gb_s", "value": "741.352", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "655.481", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "5.776", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "5.733", "unit": "GB" } ], "model": { "name": "Qwen3-4B", "hfRepo": "Qwen/Qwen3-4B", "revision": "1cfa9a7208912126459214e8b04321603b3df60c" }, "compression": { "profile": "gulp", "bitsPerWeight": "10.89", "meanTensorBitsPerWeight": "10.884", "packedBytes": 4946199508, "residentBytes": 5733001536 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-4b_nvidia-l40s_2026-10-01_gulp.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 48.305799168, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T07:40:49.975003+00:00", "version": "1.0.0", "environment": { "deviceClass": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "memoryBytes": 133680857088, "memoryKind": "unified", "cpuInfo": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "gpuInfo": "amdgpu 0x1586 vram=0.54GB gtt=133.14GB", "os": "Ubuntu 26.04.1 LTS", "engine": "drinkme 1.0.0 / torch 2.12.0a0+rocm7.13.0a20260411 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv twin / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "14.23", "unit": "tok/s", "samples": [ "14.23", "14.23", "14.23" ] }, { "name": "stock_prefill_tok_s", "value": "1040.21", "unit": "tok/s", "samples": [ "1040.21", "1036.59", "1042.13" ] }, { "name": "stock_ttft_s", "value": "0.508", "unit": "s", "samples": [ "0.5076", "0.508", "0.5126" ] }, { "name": "stock_weights_gb", "value": "16.384", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "15.136", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "14.23", "unit": "tok/s", "samples": [ "14.23", "14.22", "14.23" ] }, { "name": "twin_weights_gb", "value": "16.417", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "15.136", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "16.62", "unit": "tok/s", "samples": [ "16.59", "16.62", "16.65" ] }, { "name": "compressed_prefill_tok_s", "value": "800.94", "unit": "tok/s", "samples": [ "800.94", "802.61", "798.55" ] }, { "name": "compressed_ttft_s", "value": "0.6701", "unit": "s", "samples": [ "0.6701", "0.6688", "0.6704" ] }, { "name": "read_gb_s", "value": "229.79", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "207.574", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "12.096", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "10.761", "unit": "GB" } ], "model": { "name": "Qwen3-8B", "hfRepo": "Qwen/Qwen3-8B", "revision": "b968826d9c46dd6066d109eabc6255188de91218" }, "compression": { "profile": "sip", "bitsPerWeight": "11.374", "meanTensorBitsPerWeight": "11.346", "packedBytes": 10760735477, "residentBytes": 12006244888 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-8b_amd-ryzen-ai-max-395-w-radeon-8060s_2026-10-01.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 133.680857088, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T07:42:53.661213+00:00", "version": "1.0.0", "environment": { "deviceClass": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "memoryBytes": 133680857088, "memoryKind": "unified", "cpuInfo": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "gpuInfo": "amdgpu 0x1586 vram=0.54GB gtt=133.14GB", "os": "Ubuntu 26.04.1 LTS", "engine": "drinkme 1.0.0 / torch 2.12.0a0+rocm7.13.0a20260411 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv twin / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "14.23", "unit": "tok/s", "samples": [ "14.23", "14.23", "14.23" ] }, { "name": "stock_prefill_tok_s", "value": "1042.17", "unit": "tok/s", "samples": [ "1048.28", "1042.17", "1040.51" ] }, { "name": "stock_ttft_s", "value": "0.5103", "unit": "s", "samples": [ "0.5118", "0.5103", "0.5066" ] }, { "name": "stock_weights_gb", "value": "16.384", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "15.136", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "14.23", "unit": "tok/s", "samples": [ "14.23", "14.23", "14.21" ] }, { "name": "twin_weights_gb", "value": "16.417", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "15.136", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "15.35", "unit": "tok/s", "samples": [ "15.32", "15.35", "15.38" ] }, { "name": "compressed_prefill_tok_s", "value": "802.23", "unit": "tok/s", "samples": [ "802.2", "805.51", "802.23" ] }, { "name": "compressed_ttft_s", "value": "0.6714", "unit": "s", "samples": [ "0.6677", "0.6714", "0.6738" ] }, { "name": "read_gb_s", "value": "240.068", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "207.874", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "11.642", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "10.33", "unit": "GB" } ], "model": { "name": "Qwen3-8B", "hfRepo": "Qwen/Qwen3-8B", "revision": "b968826d9c46dd6066d109eabc6255188de91218" }, "compression": { "profile": "gulp", "bitsPerWeight": "10.903", "meanTensorBitsPerWeight": "10.88", "packedBytes": 10314829151, "residentBytes": 11575275304 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-8b_amd-ryzen-ai-max-395-w-radeon-8060s_2026-10-01_gulp.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 133.680857088, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-09-30T01:09:26.249235+00:00", "version": "1.0.0", "environment": { "deviceClass": "NVIDIA A100 80GB PCIe", "memoryBytes": 85899345920, "memoryKind": "vram", "cpuInfo": "unknown", "gpuInfo": "NVIDIA A100 80GB PCIe, 81920", "os": "Debian GNU/Linux 12 (bookworm)", "engine": "drinkme 1.0.0 / torch 2.13.0+cu130 / deltanet none / conv none / narrow-gemv triton / stock-gemv linear / raw-gemv stock / decode cudagraph", "platform": "cuda" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "66.63", "unit": "tok/s", "samples": [ "66.64", "66.63", "66.63" ] }, { "name": "stock_prefill_tok_s", "value": "9711.28", "unit": "tok/s", "samples": [ "9742.78", "9675.75", "9711.28" ] }, { "name": "stock_ttft_s", "value": "0.0548", "unit": "s", "samples": [ "0.0548", "0.0547", "0.0556" ] }, { "name": "stock_weights_gb", "value": "16.384", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "15.136", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "73.0", "unit": "tok/s", "samples": [ "73.0", "73.01", "73.0" ] }, { "name": "twin_weights_gb", "value": "16.392", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "15.136", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "46.42", "unit": "tok/s", "samples": [ "46.42", "46.42", "46.44" ] }, { "name": "compressed_prefill_tok_s", "value": "7209.18", "unit": "tok/s", "samples": [ "7135.61", "7247.74", "7209.18" ] }, { "name": "compressed_ttft_s", "value": "0.0741", "unit": "s", "samples": [ "0.0741", "0.0749", "0.0739" ] }, { "name": "read_gb_s", "value": "1800.756", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "1717.702", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "12.07", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "10.761", "unit": "GB" } ], "model": { "name": "Qwen3-8B", "hfRepo": "Qwen/Qwen3-8B", "revision": "b968826d9c46dd6066d109eabc6255188de91218" }, "compression": { "profile": "sip", "bitsPerWeight": "11.374", "meanTensorBitsPerWeight": "11.346" }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-8b_nvidia-a100-80gb-pcie_2026-09-30.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 85.89934592, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-09-30T01:09:47.591175+00:00", "version": "1.0.0", "environment": { "deviceClass": "NVIDIA H100 80GB HBM3", "memoryBytes": 85520809984, "memoryKind": "vram", "cpuInfo": "unknown", "gpuInfo": "NVIDIA H100 80GB HBM3, 81559", "os": "Debian GNU/Linux 12 (bookworm)", "engine": "drinkme 1.0.0 / torch 2.13.0+cu130 / deltanet none / conv none / narrow-gemv triton / stock-gemv linear / raw-gemv stock / decode cudagraph", "platform": "cuda" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "96.8", "unit": "tok/s", "samples": [ "96.8", "96.79", "96.86" ] }, { "name": "stock_prefill_tok_s", "value": "18454.56", "unit": "tok/s", "samples": [ "17839.21", "18454.56", "18483.15" ] }, { "name": "stock_ttft_s", "value": "0.0284", "unit": "s", "samples": [ "0.0543", "0.0284", "0.0284" ] }, { "name": "stock_weights_gb", "value": "16.384", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "15.136", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "103.29", "unit": "tok/s", "samples": [ "103.29", "103.2", "103.32" ] }, { "name": "twin_weights_gb", "value": "16.417", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "15.136", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "69.73", "unit": "tok/s", "samples": [ "69.73", "69.73", "69.61" ] }, { "name": "compressed_prefill_tok_s", "value": "10323.45", "unit": "tok/s", "samples": [ "10201.87", "10323.45", "10378.33" ] }, { "name": "compressed_ttft_s", "value": "0.0492", "unit": "s", "samples": [ "0.0927", "0.0492", "0.0472" ] }, { "name": "read_gb_s", "value": "3133.157", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "3041.175", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "12.096", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "10.761", "unit": "GB" } ], "model": { "name": "Qwen3-8B", "hfRepo": "Qwen/Qwen3-8B", "revision": "b968826d9c46dd6066d109eabc6255188de91218" }, "compression": { "profile": "sip", "bitsPerWeight": "11.374", "meanTensorBitsPerWeight": "11.346" }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-8b_nvidia-h100-80gb-hbm3_2026-09-30.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 85.520809984, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-09-30T01:10:07.254575+00:00", "version": "1.0.0", "environment": { "deviceClass": "NVIDIA L40S", "memoryBytes": 48305799168, "memoryKind": "vram", "cpuInfo": "unknown", "gpuInfo": "NVIDIA L40S, 46068", "os": "Debian GNU/Linux 12 (bookworm)", "engine": "drinkme 1.0.0 / torch 2.13.0+cu130 / deltanet none / conv none / narrow-gemv triton / stock-gemv linear / raw-gemv stock / decode cudagraph", "platform": "cuda" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "41.7", "unit": "tok/s", "samples": [ "41.68", "41.7", "41.7" ] }, { "name": "stock_prefill_tok_s", "value": "10015.59", "unit": "tok/s", "samples": [ "10162.24", "10015.59", "9831.14" ] }, { "name": "stock_ttft_s", "value": "0.0538", "unit": "s", "samples": [ "0.0607", "0.0526", "0.0538" ] }, { "name": "stock_weights_gb", "value": "16.384", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "15.136", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "42.82", "unit": "tok/s", "samples": [ "42.8", "42.82", "42.82" ] }, { "name": "twin_weights_gb", "value": "16.392", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "15.136", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "52.52", "unit": "tok/s", "samples": [ "52.52", "52.55", "52.5" ] }, { "name": "compressed_prefill_tok_s", "value": "6606.89", "unit": "tok/s", "samples": [ "6612.14", "6602.73", "6606.89" ] }, { "name": "compressed_ttft_s", "value": "0.0824", "unit": "s", "samples": [ "0.0939", "0.0823", "0.0824" ] }, { "name": "read_gb_s", "value": "741.053", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "653.498", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "12.07", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "10.761", "unit": "GB" } ], "model": { "name": "Qwen3-8B", "hfRepo": "Qwen/Qwen3-8B", "revision": "b968826d9c46dd6066d109eabc6255188de91218" }, "compression": { "profile": "sip", "bitsPerWeight": "11.374", "meanTensorBitsPerWeight": "11.346" }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-8b_nvidia-l40s_2026-09-30.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 48.305799168, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T09:59:31.795475+00:00", "version": "1.0.0", "environment": { "deviceClass": "NVIDIA L40S", "memoryBytes": 48305799168, "memoryKind": "vram", "cpuInfo": "unknown", "gpuInfo": "NVIDIA L40S, 46068", "os": "Debian GNU/Linux 12 (bookworm)", "engine": "drinkme 1.0.0 / torch 2.13.0+cu130 / deltanet none / conv none / narrow-gemv triton / stock-gemv linear / raw-gemv stock / decode cudagraph", "platform": "cuda" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "42.03", "unit": "tok/s", "samples": [ "42.03", "42.03", "42.03" ] }, { "name": "stock_prefill_tok_s", "value": "9879.62", "unit": "tok/s", "samples": [ "9912.55", "9879.62", "9847.09" ] }, { "name": "stock_ttft_s", "value": "0.0541", "unit": "s", "samples": [ "0.055", "0.054", "0.0541" ] }, { "name": "stock_weights_gb", "value": "16.384", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "15.136", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "42.77", "unit": "tok/s", "samples": [ "42.77", "42.77", "42.77" ] }, { "name": "twin_weights_gb", "value": "16.392", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "15.136", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "52.13", "unit": "tok/s", "samples": [ "52.15", "52.13", "52.1" ] }, { "name": "compressed_prefill_tok_s", "value": "6595.32", "unit": "tok/s", "samples": [ "6595.32", "6591.03", "6602.3" ] }, { "name": "compressed_ttft_s", "value": "0.0826", "unit": "s", "samples": [ "0.0839", "0.0825", "0.0826" ] }, { "name": "read_gb_s", "value": "742.299", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "655.317", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "11.617", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "10.33", "unit": "GB" } ], "model": { "name": "Qwen3-8B", "hfRepo": "Qwen/Qwen3-8B", "revision": "b968826d9c46dd6066d109eabc6255188de91218" }, "compression": { "profile": "gulp", "bitsPerWeight": "10.903", "meanTensorBitsPerWeight": "10.88", "packedBytes": 10314829151, "residentBytes": 11575275304 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-8b_nvidia-l40s_2026-10-01_gulp.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 48.305799168, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-09-30T01:10:03.923269+00:00", "version": "1.0.0", "environment": { "deviceClass": "NVIDIA L4", "memoryBytes": 24152899584, "memoryKind": "vram", "cpuInfo": "unknown", "gpuInfo": "NVIDIA L4, 23034", "os": "Debian GNU/Linux 12 (bookworm)", "engine": "drinkme 1.0.0 / torch 2.13.0+cu130 / deltanet none / conv none / narrow-gemv triton / stock-gemv linear / raw-gemv stock / decode cudagraph", "platform": "cuda" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "15.99", "unit": "tok/s", "samples": [ "15.99", "15.99", "15.98" ] }, { "name": "stock_prefill_tok_s", "value": "2626.03", "unit": "tok/s", "samples": [ "2641.8", "2626.03", "2621.88" ] }, { "name": "stock_ttft_s", "value": "0.2162", "unit": "s", "samples": [ "0.2181", "0.2162", "0.2075" ] }, { "name": "stock_weights_gb", "value": "16.384", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "15.136", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "16.26", "unit": "tok/s", "samples": [ "16.27", "16.26", "16.25" ] }, { "name": "twin_weights_gb", "value": "16.392", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "15.136", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "19.13", "unit": "tok/s", "samples": [ "19.21", "19.1", "19.13" ] }, { "name": "compressed_prefill_tok_s", "value": "1761.22", "unit": "tok/s", "samples": [ "1761.22", "1747.19", "1832.62" ] }, { "name": "compressed_ttft_s", "value": "0.3022", "unit": "s", "samples": [ "0.3123", "0.3001", "0.3022" ] }, { "name": "read_gb_s", "value": "260.645", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "231.295", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "12.07", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "10.761", "unit": "GB" } ], "model": { "name": "Qwen3-8B", "hfRepo": "Qwen/Qwen3-8B", "revision": "b968826d9c46dd6066d109eabc6255188de91218" }, "compression": { "profile": "sip", "bitsPerWeight": "11.374", "meanTensorBitsPerWeight": "11.346" }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-8b_nvidia-l4_2026-09-30.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 24.152899584, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T10:01:54.170429+00:00", "version": "1.0.0", "environment": { "deviceClass": "NVIDIA L4", "memoryBytes": 24152899584, "memoryKind": "vram", "cpuInfo": "unknown", "gpuInfo": "NVIDIA L4, 23034", "os": "Debian GNU/Linux 12 (bookworm)", "engine": "drinkme 1.0.0 / torch 2.13.0+cu130 / deltanet none / conv none / narrow-gemv triton / stock-gemv linear / raw-gemv stock / decode cudagraph", "platform": "cuda" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "16.03", "unit": "tok/s", "samples": [ "16.04", "16.03", "16.03" ] }, { "name": "stock_prefill_tok_s", "value": "2759.58", "unit": "tok/s", "samples": [ "2726.69", "2759.58", "2772.58" ] }, { "name": "stock_ttft_s", "value": "0.199", "unit": "s", "samples": [ "0.2082", "0.199", "0.198" ] }, { "name": "stock_weights_gb", "value": "16.384", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "15.136", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "16.17", "unit": "tok/s", "samples": [ "16.18", "16.17", "16.16" ] }, { "name": "twin_weights_gb", "value": "16.392", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "15.136", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "14.85", "unit": "tok/s", "samples": [ "14.93", "14.85", "14.66" ] }, { "name": "compressed_prefill_tok_s", "value": "1798.24", "unit": "tok/s", "samples": [ "1781.55", "1800.41", "1798.24" ] }, { "name": "compressed_ttft_s", "value": "0.3035", "unit": "s", "samples": [ "0.3035", "0.2993", "0.3044" ] }, { "name": "read_gb_s", "value": "260.43", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "231.035", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "11.617", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "10.33", "unit": "GB" } ], "model": { "name": "Qwen3-8B", "hfRepo": "Qwen/Qwen3-8B", "revision": "b968826d9c46dd6066d109eabc6255188de91218" }, "compression": { "profile": "gulp", "bitsPerWeight": "10.903", "meanTensorBitsPerWeight": "10.88", "packedBytes": 10314829151, "residentBytes": 11575275304 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-8b_nvidia-l4_2026-10-01_gulp.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 24.152899584, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T07:50:46.935516+00:00", "version": "1.0.0", "environment": { "deviceClass": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "memoryBytes": 133680857088, "memoryKind": "unified", "cpuInfo": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "gpuInfo": "amdgpu 0x1586 vram=0.54GB gtt=133.14GB", "os": "Ubuntu 26.04.1 LTS", "engine": "drinkme 1.0.0 / torch 2.12.0a0+rocm7.13.0a20260411 / deltanet fla 0.5.2 / conv fla 0.5.2 / narrow-gemv triton / stock-gemv triton / raw-gemv twin / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "4.24", "unit": "tok/s", "samples": [ "4.24", "4.24", "4.24" ] }, { "name": "stock_prefill_tok_s", "value": "461.22", "unit": "tok/s", "samples": [ "460.65", "461.22", "461.84" ] }, { "name": "stock_ttft_s", "value": "1.1417", "unit": "s", "samples": [ "1.1417", "1.1426", "1.1384" ] }, { "name": "stock_weights_gb", "value": "54.729", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "51.244", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "4.25", "unit": "tok/s", "samples": [ "4.25", "4.25", "4.25" ] }, { "name": "twin_weights_gb", "value": "54.75", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "51.244", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "5.01", "unit": "tok/s", "samples": [ "5.01", "5.01", "5.01" ] }, { "name": "compressed_prefill_tok_s", "value": "332.09", "unit": "tok/s", "samples": [ "331.74", "332.09", "332.28" ] }, { "name": "compressed_ttft_s", "value": "1.5895", "unit": "s", "samples": [ "1.6122", "1.5895", "1.5894" ] }, { "name": "read_gb_s", "value": "238.519", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "202.493", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "39.405", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "36.044", "unit": "GB" }, { "name": "stock_spec_decode_tok_s_agent", "value": "7.61", "unit": "tok/s", "samples": [ "7.6", "7.61", "7.61" ] }, { "name": "stock_spec_decode_tok_s_chat", "value": "8.23", "unit": "tok/s", "samples": [ "8.23", "8.23", "8.23" ] }, { "name": "stock_spec_decode_tok_s_code", "value": "10.12", "unit": "tok/s", "samples": [ "10.12", "10.12", "10.12" ] }, { "name": "compressed_spec_decode_tok_s_agent", "value": "10.11", "unit": "tok/s", "samples": [ "10.12", "10.11", "10.11" ] }, { "name": "compressed_spec_decode_tok_s_chat", "value": "10.83", "unit": "tok/s", "samples": [ "10.83", "10.81", "10.83" ] }, { "name": "compressed_spec_decode_tok_s_code", "value": "13.12", "unit": "tok/s", "samples": [ "13.23", "13.12", "13.08" ] } ], "model": { "name": "Qwen3.8-27B", "hfRepo": "Qwen/Qwen3.8-27B", "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0" }, "compression": { "profile": "sip", "bitsPerWeight": "11.25", "meanTensorBitsPerWeight": "11.264", "packedBytes": 36640893751, "residentBytes": 39246304524 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3.8-27b_amd-ryzen-ai-max-395-w-radeon-8060s_2026-10-01.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "stock_spec_decode_tok_s_agent": "/metrics/15/value", "stock_spec_decode_tok_s_chat": "/metrics/16/value", "stock_spec_decode_tok_s_code": "/metrics/17/value", "compressed_spec_decode_tok_s_agent": "/metrics/18/value", "compressed_spec_decode_tok_s_chat": "/metrics/19/value", "compressed_spec_decode_tok_s_code": "/metrics/20/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 133.680857088, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T08:08:55.222438+00:00", "version": "1.0.0", "environment": { "deviceClass": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "memoryBytes": 133680857088, "memoryKind": "unified", "cpuInfo": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S", "gpuInfo": "amdgpu 0x1586 vram=0.54GB gtt=133.14GB", "os": "Ubuntu 26.04.1 LTS", "engine": "drinkme 1.0.0 / torch 2.12.0a0+rocm7.13.0a20260411 / deltanet fla 0.5.2 / conv fla 0.5.2 / narrow-gemv triton / stock-gemv triton / raw-gemv twin / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "4.25", "unit": "tok/s", "samples": [ "4.25", "4.25", "4.25" ] }, { "name": "stock_prefill_tok_s", "value": "463.33", "unit": "tok/s", "samples": [ "462.34", "463.33", "463.41" ] }, { "name": "stock_ttft_s", "value": "1.1362", "unit": "s", "samples": [ "1.1368", "1.1333", "1.1362" ] }, { "name": "stock_weights_gb", "value": "54.729", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "51.244", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "4.23", "unit": "tok/s", "samples": [ "4.23", "4.15", "4.23" ] }, { "name": "twin_weights_gb", "value": "54.75", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "51.244", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "4.69", "unit": "tok/s", "samples": [ "4.69", "4.69", "4.7" ] }, { "name": "compressed_prefill_tok_s", "value": "330.41", "unit": "tok/s", "samples": [ "330.59", "330.41", "329.74" ] }, { "name": "compressed_ttft_s", "value": "1.5979", "unit": "s", "samples": [ "1.5979", "1.5996", "1.5933" ] }, { "name": "read_gb_s", "value": "239.738", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "207.984", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "37.885", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "34.622", "unit": "GB" }, { "name": "stock_spec_decode_tok_s_agent", "value": "7.7", "unit": "tok/s", "samples": [ "7.7", "7.7", "7.7" ] }, { "name": "stock_spec_decode_tok_s_chat", "value": "8.33", "unit": "tok/s", "samples": [ "8.34", "8.33", "8.33" ] }, { "name": "stock_spec_decode_tok_s_code", "value": "10.21", "unit": "tok/s", "samples": [ "10.23", "10.2", "10.21" ] }, { "name": "compressed_spec_decode_tok_s_agent", "value": "9.01", "unit": "tok/s", "samples": [ "9.02", "9.01", "9.0" ] }, { "name": "compressed_spec_decode_tok_s_chat", "value": "9.39", "unit": "tok/s", "samples": [ "9.55", "9.39", "9.34" ] }, { "name": "compressed_spec_decode_tok_s_code", "value": "11.61", "unit": "tok/s", "samples": [ "11.64", "11.61", "11.61" ] } ], "model": { "name": "Qwen3.8-27B", "hfRepo": "Qwen/Qwen3.8-27B", "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0" }, "compression": { "profile": "gulp", "bitsPerWeight": "10.791", "meanTensorBitsPerWeight": "10.803", "packedBytes": 35143172017, "residentBytes": 37800264088 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3.8-27b_amd-ryzen-ai-max-395-w-radeon-8060s_2026-10-01_gulp.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "stock_spec_decode_tok_s_agent": "/metrics/15/value", "stock_spec_decode_tok_s_chat": "/metrics/16/value", "stock_spec_decode_tok_s_code": "/metrics/17/value", "compressed_spec_decode_tok_s_agent": "/metrics/18/value", "compressed_spec_decode_tok_s_chat": "/metrics/19/value", "compressed_spec_decode_tok_s_code": "/metrics/20/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 133.680857088, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-09-30T01:14:11.972020+00:00", "version": "1.0.0", "environment": { "deviceClass": "NVIDIA A100-SXM4-80GB", "memoryBytes": 85899345920, "memoryKind": "vram", "cpuInfo": "unknown", "gpuInfo": "NVIDIA A100-SXM4-80GB, 81920", "os": "Debian GNU/Linux 12 (bookworm)", "engine": "drinkme 1.0.0 / torch 2.13.0+cu130 / deltanet fla 0.5.2 / conv torch / narrow-gemv triton / stock-gemv linear / raw-gemv stock / decode cudagraph", "platform": "cuda" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "23.38", "unit": "tok/s", "samples": [ "23.39", "23.38", "23.38" ] }, { "name": "stock_prefill_tok_s", "value": "3170.94", "unit": "tok/s", "samples": [ "3170.26", "3170.94", "3172.04" ] }, { "name": "stock_ttft_s", "value": "0.1652", "unit": "s", "samples": [ "0.1822", "0.1652", "0.165" ] }, { "name": "stock_weights_gb", "value": "54.729", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "51.244", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "24.89", "unit": "tok/s", "samples": [ "24.9", "24.89", "24.88" ] }, { "name": "twin_weights_gb", "value": "54.725", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "51.244", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "15.82", "unit": "tok/s", "samples": [ "15.8", "15.82", "15.85" ] }, { "name": "compressed_prefill_tok_s", "value": "2342.25", "unit": "tok/s", "samples": [ "2363.14", "2339.77", "2342.25" ] }, { "name": "compressed_ttft_s", "value": "0.224", "unit": "s", "samples": [ "0.2273", "0.224", "0.2236" ] }, { "name": "read_gb_s", "value": "1731.521", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "1766.083", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "39.382", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "36.044", "unit": "GB" }, { "name": "stock_spec_decode_tok_s_agent", "value": "45.19", "unit": "tok/s", "samples": [ "45.33", "44.71", "45.19" ] }, { "name": "stock_spec_decode_tok_s_chat", "value": "48.13", "unit": "tok/s", "samples": [ "48.13", "48.26", "48.1" ] }, { "name": "stock_spec_decode_tok_s_code", "value": "62.57", "unit": "tok/s", "samples": [ "62.49", "62.57", "62.59" ] }, { "name": "compressed_spec_decode_tok_s_agent", "value": "30.49", "unit": "tok/s", "samples": [ "30.45", "30.53", "30.49" ] }, { "name": "compressed_spec_decode_tok_s_chat", "value": "33.56", "unit": "tok/s", "samples": [ "33.59", "33.56", "33.5" ] }, { "name": "compressed_spec_decode_tok_s_code", "value": "40.13", "unit": "tok/s", "samples": [ "40.15", "40.13", "40.1" ] } ], "model": { "name": "Qwen3.8-27B", "hfRepo": "Qwen/Qwen3.8-27B", "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0" }, "compression": { "profile": "sip", "bitsPerWeight": "11.25", "meanTensorBitsPerWeight": "11.264" }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3.8-27b_nvidia-a100-sxm4-80gb_2026-09-30.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "stock_spec_decode_tok_s_agent": "/metrics/15/value", "stock_spec_decode_tok_s_chat": "/metrics/16/value", "stock_spec_decode_tok_s_code": "/metrics/17/value", "compressed_spec_decode_tok_s_agent": "/metrics/18/value", "compressed_spec_decode_tok_s_chat": "/metrics/19/value", "compressed_spec_decode_tok_s_code": "/metrics/20/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 85.89934592, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-09-30T01:15:00.714139+00:00", "version": "1.0.0", "environment": { "deviceClass": "NVIDIA H100 80GB HBM3", "memoryBytes": 85520809984, "memoryKind": "vram", "cpuInfo": "unknown", "gpuInfo": "NVIDIA H100 80GB HBM3, 81559", "os": "Debian GNU/Linux 12 (bookworm)", "engine": "drinkme 1.0.0 / torch 2.13.0+cu130 / deltanet fla 0.5.2 / conv torch / narrow-gemv triton / stock-gemv linear / raw-gemv stock / decode cudagraph", "platform": "cuda" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "35.39", "unit": "tok/s", "samples": [ "35.35", "35.41", "35.39" ] }, { "name": "stock_prefill_tok_s", "value": "5647.33", "unit": "tok/s", "samples": [ "5535.59", "5647.33", "5690.9" ] }, { "name": "stock_ttft_s", "value": "0.0923", "unit": "s", "samples": [ "0.154", "0.0923", "0.0919" ] }, { "name": "stock_weights_gb", "value": "54.729", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "51.244", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "37.28", "unit": "tok/s", "samples": [ "37.01", "37.4", "37.28" ] }, { "name": "twin_weights_gb", "value": "54.75", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "51.244", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "24.06", "unit": "tok/s", "samples": [ "24.05", "24.06", "24.06" ] }, { "name": "compressed_prefill_tok_s", "value": "3794.3", "unit": "tok/s", "samples": [ "3794.3", "3797.54", "3780.14" ] }, { "name": "compressed_ttft_s", "value": "0.1415", "unit": "s", "samples": [ "0.1591", "0.1415", "0.1399" ] }, { "name": "read_gb_s", "value": "3089.312", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "3025.444", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "39.406", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "36.044", "unit": "GB" }, { "name": "stock_spec_decode_tok_s_agent", "value": "70.96", "unit": "tok/s", "samples": [ "71.02", "70.96", "70.92" ] }, { "name": "stock_spec_decode_tok_s_chat", "value": "80.67", "unit": "tok/s", "samples": [ "80.67", "80.8", "80.6" ] }, { "name": "stock_spec_decode_tok_s_code", "value": "97.64", "unit": "tok/s", "samples": [ "97.64", "97.58", "97.65" ] }, { "name": "compressed_spec_decode_tok_s_agent", "value": "46.72", "unit": "tok/s", "samples": [ "46.86", "46.14", "46.72" ] }, { "name": "compressed_spec_decode_tok_s_chat", "value": "51.64", "unit": "tok/s", "samples": [ "51.64", "51.57", "51.83" ] }, { "name": "compressed_spec_decode_tok_s_code", "value": "63.13", "unit": "tok/s", "samples": [ "63.1", "63.13", "63.26" ] } ], "model": { "name": "Qwen3.8-27B", "hfRepo": "Qwen/Qwen3.8-27B", "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0" }, "compression": { "profile": "sip", "bitsPerWeight": "11.25", "meanTensorBitsPerWeight": "11.264" }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3.8-27b_nvidia-h100-80gb-hbm3_2026-09-30.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "stock_spec_decode_tok_s_agent": "/metrics/15/value", "stock_spec_decode_tok_s_chat": "/metrics/16/value", "stock_spec_decode_tok_s_code": "/metrics/17/value", "compressed_spec_decode_tok_s_agent": "/metrics/18/value", "compressed_spec_decode_tok_s_chat": "/metrics/19/value", "compressed_spec_decode_tok_s_code": "/metrics/20/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 85.520809984, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-09-30T01:14:51.668004+00:00", "version": "1.0.0", "environment": { "deviceClass": "NVIDIA L40S", "memoryBytes": 48305799168, "memoryKind": "vram", "cpuInfo": "unknown", "gpuInfo": "NVIDIA L40S, 46068", "os": "Debian GNU/Linux 12 (bookworm)", "engine": "drinkme 1.0.0 / torch 2.13.0+cu130 / deltanet fla 0.5.2 / conv torch / narrow-gemv triton / stock-gemv linear / raw-gemv stock / decode cudagraph", "platform": "cuda" }, "metrics": [ { "name": "compressed_decode_tok_s", "value": "16.61", "unit": "tok/s", "samples": [ "16.61", "16.61", "16.61" ] }, { "name": "compressed_prefill_tok_s", "value": "1833.0", "unit": "tok/s", "samples": [ "1832.87", "1833.0", "1833.89" ] }, { "name": "compressed_ttft_s", "value": "0.29", "unit": "s", "samples": [ "0.2899", "0.29", "0.2903" ] }, { "name": "read_gb_s", "value": "741.968", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "651.895", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "39.372", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "36.044", "unit": "GB" }, { "name": "compressed_spec_decode_tok_s_agent", "value": "35.49", "unit": "tok/s", "samples": [ "35.55", "35.49", "35.47" ] }, { "name": "compressed_spec_decode_tok_s_chat", "value": "39.62", "unit": "tok/s", "samples": [ "39.75", "39.62", "39.61" ] }, { "name": "compressed_spec_decode_tok_s_code", "value": "48.62", "unit": "tok/s", "samples": [ "48.62", "48.64", "48.57" ] } ], "model": { "name": "Qwen3.8-27B", "hfRepo": "Qwen/Qwen3.8-27B", "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0" }, "compression": { "profile": "sip", "bitsPerWeight": "11.25", "meanTensorBitsPerWeight": "11.264" }, "stock": { "outcome": "skipped_predicted_nonfit", "budgetBytes": 48305799168 }, "site": { "source": "measurements/qwen3.8-27b_nvidia-l40s_2026-09-30.json", "at": { "compressed_decode_tok_s": "/metrics/0/value", "compressed_prefill_tok_s": "/metrics/1/value", "compressed_ttft_s": "/metrics/2/value", "read_gb_s": "/metrics/3/value", "copy_gb_s": "/metrics/4/value", "compressed_weights_gb": "/metrics/5/value", "compressed_decode_read_gb": "/metrics/6/value", "compressed_spec_decode_tok_s_agent": "/metrics/7/value", "compressed_spec_decode_tok_s_chat": "/metrics/8/value", "compressed_spec_decode_tok_s_code": "/metrics/9/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "stock.budgetBytes": "/stock/budgetBytes", "stockError": "/raw/stock_error", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 48.305799168, "harness": "drinkme bench", "stockError": "skipped by bench's stock-arm fit check (predicted OOM under the stream loader's transient)" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-01T10:16:08.044542+00:00", "version": "1.0.0", "environment": { "deviceClass": "NVIDIA L40S", "memoryBytes": 51527024640, "memoryKind": "vram", "cpuInfo": "unknown", "gpuInfo": "NVIDIA L40S, 49140", "os": "Debian GNU/Linux 12 (bookworm)", "engine": "drinkme 1.0.0 / torch 2.13.0+cu130 / deltanet fla 0.5.2 / conv torch / narrow-gemv triton / stock-gemv linear / raw-gemv stock / decode cudagraph", "platform": "cuda" }, "metrics": [ { "name": "compressed_decode_tok_s", "value": "17.9", "unit": "tok/s", "samples": [ "17.9", "17.92", "17.9" ] }, { "name": "compressed_prefill_tok_s", "value": "2000.15", "unit": "tok/s", "samples": [ "1997.9", "2000.15", "2005.55" ] }, { "name": "compressed_ttft_s", "value": "0.2655", "unit": "s", "samples": [ "0.2656", "0.2655", "0.2643" ] }, { "name": "read_gb_s", "value": "839.097", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "766.72", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "37.852", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "34.622", "unit": "GB" }, { "name": "compressed_spec_decode_tok_s_agent", "value": "32.05", "unit": "tok/s", "samples": [ "32.29", "32.05", "31.88" ] }, { "name": "compressed_spec_decode_tok_s_chat", "value": "35.1", "unit": "tok/s", "samples": [ "35.47", "35.1", "35.0" ] }, { "name": "compressed_spec_decode_tok_s_code", "value": "42.88", "unit": "tok/s", "samples": [ "42.88", "42.97", "42.88" ] } ], "model": { "name": "Qwen3.8-27B", "hfRepo": "Qwen/Qwen3.8-27B", "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0" }, "compression": { "profile": "gulp", "bitsPerWeight": "10.791", "meanTensorBitsPerWeight": "10.803", "packedBytes": 35143172017, "residentBytes": 37800264088 }, "stock": { "outcome": "skipped_predicted_nonfit", "budgetBytes": 51527024640 }, "site": { "source": "measurements/qwen3.8-27b_nvidia-l40s_2026-10-01_gulp.json", "at": { "compressed_decode_tok_s": "/metrics/0/value", "compressed_prefill_tok_s": "/metrics/1/value", "compressed_ttft_s": "/metrics/2/value", "read_gb_s": "/metrics/3/value", "copy_gb_s": "/metrics/4/value", "compressed_weights_gb": "/metrics/5/value", "compressed_decode_read_gb": "/metrics/6/value", "compressed_spec_decode_tok_s_agent": "/metrics/7/value", "compressed_spec_decode_tok_s_chat": "/metrics/8/value", "compressed_spec_decode_tok_s_code": "/metrics/9/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "stock.budgetBytes": "/stock/budgetBytes", "stockError": "/raw/stock_error", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 51.52702464, "harness": "drinkme bench", "stockError": "skipped by bench's stock-arm fit check (predicted OOM under the stream loader's transient)" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-02T02:17:44.885686+00:00", "version": "1.0.0", "environment": { "deviceClass": "Radeon RX 7600 XT", "memoryBytes": 17163091968, "memoryKind": "vram", "cpuInfo": "AMD Ryzen 7 5700X 8-Core Processor", "gpuInfo": "amdgpu 0x7480 vram=17.16GB gtt=16.78GB", "os": "CachyOS", "engine": "drinkme 1.0.0 / torch 2.13.0+rocm7.1 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv stock / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "42.15", "unit": "tok/s", "samples": [ "42.06", "42.15", "42.21" ] }, { "name": "stock_prefill_tok_s", "value": "10444.96", "unit": "tok/s", "samples": [ "10463.94", "10420.28", "10444.96" ] }, { "name": "stock_ttft_s", "value": "0.0542", "unit": "s", "samples": [ "0.0542", "0.0542", "0.0542" ] }, { "name": "stock_weights_gb", "value": "1.192", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "1.192", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "41.85", "unit": "tok/s", "samples": [ "41.85", "41.84", "41.86" ] }, { "name": "twin_weights_gb", "value": "1.226", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "1.192", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "41.1", "unit": "tok/s", "samples": [ "41.15", "40.92", "41.1" ] }, { "name": "compressed_prefill_tok_s", "value": "10613.14", "unit": "tok/s", "samples": [ "10622.59", "10613.14", "10612.25" ] }, { "name": "compressed_ttft_s", "value": "0.0538", "unit": "s", "samples": [ "0.0533", "0.054", "0.0538" ] }, { "name": "read_gb_s", "value": "278.702", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "250.017", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "0.973", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "0.934", "unit": "GB" } ], "model": { "name": "Qwen3-0.6B", "hfRepo": "Qwen/Qwen3-0.6B", "revision": "c1899de289a04d12100db370d81485cdf75e47ca" }, "compression": { "profile": "sip", "bitsPerWeight": "11.31", "meanTensorBitsPerWeight": "11.316", "packedBytes": 622929168, "residentBytes": 1245571120 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-0.6b_radeon-rx-7600-xt_2026-10-02_rerun.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 17.163091968, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-02T01:19:23.216482+00:00", "version": "1.0.0", "environment": { "deviceClass": "Radeon RX 7600 XT", "memoryBytes": 17163091968, "memoryKind": "vram", "cpuInfo": "AMD Ryzen 7 5700X 8-Core Processor", "gpuInfo": "amdgpu 0x7480 vram=17.16GB gtt=16.78GB", "os": "CachyOS", "engine": "drinkme 1.0.0 / torch 2.13.0+rocm7.1 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv stock / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "41.99", "unit": "tok/s", "samples": [ "42.05", "41.99", "41.95" ] }, { "name": "stock_prefill_tok_s", "value": "10465.42", "unit": "tok/s", "samples": [ "10420.43", "10465.42", "10476.14" ] }, { "name": "stock_ttft_s", "value": "0.0542", "unit": "s", "samples": [ "0.0541", "0.0542", "0.0543" ] }, { "name": "stock_weights_gb", "value": "1.192", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "1.192", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "39.46", "unit": "tok/s", "samples": [ "39.54", "39.43", "39.46" ] }, { "name": "twin_weights_gb", "value": "1.226", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "1.192", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "38.95", "unit": "tok/s", "samples": [ "38.89", "38.95", "39.06" ] }, { "name": "compressed_prefill_tok_s", "value": "10699.81", "unit": "tok/s", "samples": [ "10699.49", "10699.81", "10702.9" ] }, { "name": "compressed_ttft_s", "value": "0.0536", "unit": "s", "samples": [ "0.0534", "0.0538", "0.0536" ] }, { "name": "read_gb_s", "value": "278.822", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "250.919", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "0.948", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "0.911", "unit": "GB" } ], "model": { "name": "Qwen3-0.6B", "hfRepo": "Qwen/Qwen3-0.6B", "revision": "c1899de289a04d12100db370d81485cdf75e47ca" }, "compression": { "profile": "gulp", "bitsPerWeight": "10.874", "meanTensorBitsPerWeight": "10.886", "packedBytes": 598928188, "residentBytes": 1222550636 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-0.6b_radeon-rx-7600-xt_2026-10-02_gulp.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 17.163091968, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-02T01:20:56.636904+00:00", "version": "1.0.0", "environment": { "deviceClass": "Radeon RX 7600 XT", "memoryBytes": 17163091968, "memoryKind": "vram", "cpuInfo": "AMD Ryzen 7 5700X 8-Core Processor", "gpuInfo": "amdgpu 0x7480 vram=17.16GB gtt=16.78GB", "os": "CachyOS", "engine": "drinkme 1.0.0 / torch 2.13.0+rocm7.1 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv stock / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "35.88", "unit": "tok/s", "samples": [ "35.88", "35.82", "36.12" ] }, { "name": "stock_prefill_tok_s", "value": "5576.95", "unit": "tok/s", "samples": [ "5613.66", "5576.95", "5539.68" ] }, { "name": "stock_ttft_s", "value": "0.1003", "unit": "s", "samples": [ "0.1003", "0.1012", "0.1001" ] }, { "name": "stock_weights_gb", "value": "3.442", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "3.441", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "36.15", "unit": "tok/s", "samples": [ "36.16", "35.83", "36.15" ] }, { "name": "twin_weights_gb", "value": "3.475", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "3.441", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "34.87", "unit": "tok/s", "samples": [ "34.87", "34.78", "35.22" ] }, { "name": "compressed_prefill_tok_s", "value": "4961.18", "unit": "tok/s", "samples": [ "4961.18", "4960.69", "4963.08" ] }, { "name": "compressed_ttft_s", "value": "0.112", "unit": "s", "samples": [ "0.112", "0.112", "0.1119" ] }, { "name": "read_gb_s", "value": "278.581", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "250.213", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "2.647", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "2.607", "unit": "GB" } ], "model": { "name": "Qwen3-1.7B", "hfRepo": "Qwen/Qwen3-1.7B", "revision": "70d244cc86ccca08cf5af4e1e306ecf908b1ad5e" }, "compression": { "profile": "sip", "bitsPerWeight": "11.263", "meanTensorBitsPerWeight": "11.273", "packedBytes": 1984349124, "residentBytes": 3229437676 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-1.7b_radeon-rx-7600-xt_2026-10-02.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 17.163091968, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-02T01:23:27.108633+00:00", "version": "1.0.0", "environment": { "deviceClass": "Radeon RX 7600 XT", "memoryBytes": 17163091968, "memoryKind": "vram", "cpuInfo": "AMD Ryzen 7 5700X 8-Core Processor", "gpuInfo": "amdgpu 0x7480 vram=17.16GB gtt=16.78GB", "os": "CachyOS", "engine": "drinkme 1.0.0 / torch 2.13.0+rocm7.1 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv stock / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "36.92", "unit": "tok/s", "samples": [ "36.67", "36.94", "36.92" ] }, { "name": "stock_prefill_tok_s", "value": "5576.58", "unit": "tok/s", "samples": [ "5576.79", "5545.81", "5576.58" ] }, { "name": "stock_ttft_s", "value": "0.1002", "unit": "s", "samples": [ "0.1002", "0.1009", "0.1002" ] }, { "name": "stock_weights_gb", "value": "3.442", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "3.441", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "35.57", "unit": "tok/s", "samples": [ "35.1", "35.66", "35.57" ] }, { "name": "twin_weights_gb", "value": "3.475", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "3.441", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "34.49", "unit": "tok/s", "samples": [ "35.1", "34.49", "34.23" ] }, { "name": "compressed_prefill_tok_s", "value": "4908.94", "unit": "tok/s", "samples": [ "4902.84", "4909.34", "4908.94" ] }, { "name": "compressed_ttft_s", "value": "0.113", "unit": "s", "samples": [ "0.113", "0.1129", "0.1132" ] }, { "name": "read_gb_s", "value": "278.654", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "250.04", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "2.571", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "2.531", "unit": "GB" } ], "model": { "name": "Qwen3-1.7B", "hfRepo": "Qwen/Qwen3-1.7B", "revision": "70d244cc86ccca08cf5af4e1e306ecf908b1ad5e" }, "compression": { "profile": "gulp", "bitsPerWeight": "10.816", "meanTensorBitsPerWeight": "10.833", "packedBytes": 1905699240, "residentBytes": 3153660684 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-1.7b_radeon-rx-7600-xt_2026-10-02_gulp.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 17.163091968, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-02T01:24:55.824796+00:00", "version": "1.0.0", "environment": { "deviceClass": "Radeon RX 7600 XT", "memoryBytes": 17163091968, "memoryKind": "vram", "cpuInfo": "AMD Ryzen 7 5700X 8-Core Processor", "gpuInfo": "amdgpu 0x7480 vram=17.16GB gtt=16.78GB", "os": "CachyOS", "engine": "drinkme 1.0.0 / torch 2.13.0+rocm7.1 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv stock / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "23.44", "unit": "tok/s", "samples": [ "23.44", "23.45", "23.18" ] }, { "name": "stock_prefill_tok_s", "value": "2790.96", "unit": "tok/s", "samples": [ "2791.63", "2788.88", "2790.96" ] }, { "name": "stock_ttft_s", "value": "0.1942", "unit": "s", "samples": [ "0.1938", "0.1942", "0.1942" ] }, { "name": "stock_weights_gb", "value": "8.102", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "8.045", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "23.19", "unit": "tok/s", "samples": [ "23.57", "23.19", "23.19" ] }, { "name": "twin_weights_gb", "value": "8.135", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "8.045", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "19.83", "unit": "tok/s", "samples": [ "19.84", "19.83", "19.83" ] }, { "name": "compressed_prefill_tok_s", "value": "2200.23", "unit": "tok/s", "samples": [ "2189.54", "2200.23", "2200.59" ] }, { "name": "compressed_ttft_s", "value": "0.2428", "unit": "s", "samples": [ "0.2428", "0.2433", "0.2423" ] }, { "name": "read_gb_s", "value": "278.803", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "250.15", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "6.02", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "5.92", "unit": "GB" } ], "model": { "name": "Qwen3-4B", "hfRepo": "Qwen/Qwen3-4B", "revision": "1cfa9a7208912126459214e8b04321603b3df60c" }, "compression": { "profile": "sip", "bitsPerWeight": "11.32", "meanTensorBitsPerWeight": "11.314", "packedBytes": 5141508116, "residentBytes": 5920045376 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-4b_radeon-rx-7600-xt_2026-10-02.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 17.163091968, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-02T01:28:47.225065+00:00", "version": "1.0.0", "environment": { "deviceClass": "Radeon RX 7600 XT", "memoryBytes": 17163091968, "memoryKind": "vram", "cpuInfo": "AMD Ryzen 7 5700X 8-Core Processor", "gpuInfo": "amdgpu 0x7480 vram=17.16GB gtt=16.78GB", "os": "CachyOS", "engine": "drinkme 1.0.0 / torch 2.13.0+rocm7.1 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv stock / decode eager", "platform": "rocm" }, "metrics": [ { "name": "stock_decode_tok_s", "value": "23.45", "unit": "tok/s", "samples": [ "23.45", "23.45", "23.2" ] }, { "name": "stock_prefill_tok_s", "value": "2787.29", "unit": "tok/s", "samples": [ "2792.62", "2787.29", "2786.79" ] }, { "name": "stock_ttft_s", "value": "0.194", "unit": "s", "samples": [ "0.1939", "0.194", "0.194" ] }, { "name": "stock_weights_gb", "value": "8.102", "unit": "GB" }, { "name": "stock_decode_read_gb", "value": "8.045", "unit": "GB" }, { "name": "twin_decode_tok_s", "value": "23.71", "unit": "tok/s", "samples": [ "23.66", "23.72", "23.71" ] }, { "name": "twin_weights_gb", "value": "8.135", "unit": "GB" }, { "name": "twin_decode_read_gb", "value": "8.045", "unit": "GB" }, { "name": "compressed_decode_tok_s", "value": "12.82", "unit": "tok/s", "samples": [ "12.8", "12.82", "12.85" ] }, { "name": "compressed_prefill_tok_s", "value": "2075.85", "unit": "tok/s", "samples": [ "2076.56", "2075.85", "2074.61" ] }, { "name": "compressed_ttft_s", "value": "0.2586", "unit": "s", "samples": [ "0.2574", "0.2588", "0.2586" ] }, { "name": "read_gb_s", "value": "278.696", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "250.267", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "5.8", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "5.733", "unit": "GB" } ], "model": { "name": "Qwen3-4B", "hfRepo": "Qwen/Qwen3-4B", "revision": "1cfa9a7208912126459214e8b04321603b3df60c" }, "compression": { "profile": "gulp", "bitsPerWeight": "10.89", "meanTensorBitsPerWeight": "10.884", "packedBytes": 4946199508, "residentBytes": 5733001536 }, "stock": { "outcome": "measured" }, "site": { "source": "measurements/qwen3-4b_radeon-rx-7600-xt_2026-10-02_gulp.json", "at": { "stock_decode_tok_s": "/metrics/0/value", "stock_prefill_tok_s": "/metrics/1/value", "stock_ttft_s": "/metrics/2/value", "stock_weights_gb": "/metrics/3/value", "stock_decode_read_gb": "/metrics/4/value", "twin_decode_tok_s": "/metrics/5/value", "twin_weights_gb": "/metrics/6/value", "twin_decode_read_gb": "/metrics/7/value", "compressed_decode_tok_s": "/metrics/8/value", "compressed_prefill_tok_s": "/metrics/9/value", "compressed_ttft_s": "/metrics/10/value", "read_gb_s": "/metrics/11/value", "copy_gb_s": "/metrics/12/value", "compressed_weights_gb": "/metrics/13/value", "compressed_decode_read_gb": "/metrics/14/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 17.163091968, "harness": "drinkme bench" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-02T01:32:41.313458+00:00", "version": "1.0.0", "environment": { "deviceClass": "Radeon RX 7600 XT", "memoryBytes": 17163091968, "memoryKind": "vram", "cpuInfo": "AMD Ryzen 7 5700X 8-Core Processor", "gpuInfo": "amdgpu 0x7480 vram=17.16GB gtt=16.78GB", "os": "CachyOS", "engine": "drinkme 1.0.0 / torch 2.13.0+rocm7.1 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv stock / decode eager", "platform": "rocm" }, "metrics": [ { "name": "compressed_decode_tok_s", "value": "13.47", "unit": "tok/s", "samples": [ "13.54", "13.46", "13.47" ] }, { "name": "compressed_prefill_tok_s", "value": "1136.65", "unit": "tok/s", "samples": [ "1136.65", "1138.82", "1136.24" ] }, { "name": "compressed_ttft_s", "value": "0.4758", "unit": "s", "samples": [ "0.4755", "0.4758", "0.4765" ] }, { "name": "read_gb_s", "value": "278.669", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "250.134", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "12.062", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "10.761", "unit": "GB" } ], "model": { "name": "Qwen3-8B", "hfRepo": "Qwen/Qwen3-8B", "revision": "b968826d9c46dd6066d109eabc6255188de91218" }, "compression": { "profile": "sip", "bitsPerWeight": "11.374", "meanTensorBitsPerWeight": "11.346", "packedBytes": 10760735477, "residentBytes": 12006244888 }, "stock": { "outcome": "skipped_predicted_nonfit", "budgetBytes": 17163091968 }, "site": { "source": "measurements/qwen3-8b_radeon-rx-7600-xt_2026-10-02.json", "at": { "compressed_decode_tok_s": "/metrics/0/value", "compressed_prefill_tok_s": "/metrics/1/value", "compressed_ttft_s": "/metrics/2/value", "read_gb_s": "/metrics/3/value", "copy_gb_s": "/metrics/4/value", "compressed_weights_gb": "/metrics/5/value", "compressed_decode_read_gb": "/metrics/6/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "stock.budgetBytes": "/stock/budgetBytes", "stockError": "/raw/stock_error", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 17.163091968, "harness": "drinkme bench", "stockError": "skipped by bench's stock-arm fit check (predicted OOM under the stream loader's transient)" } }, { "$type": "wtf.petrichor.drinkme.measurement", "createdAt": "2026-10-02T01:39:00.533121+00:00", "version": "1.0.0", "environment": { "deviceClass": "Radeon RX 7600 XT", "memoryBytes": 17163091968, "memoryKind": "vram", "cpuInfo": "AMD Ryzen 7 5700X 8-Core Processor", "gpuInfo": "amdgpu 0x7480 vram=17.16GB gtt=16.78GB", "os": "CachyOS", "engine": "drinkme 1.0.0 / torch 2.13.0+rocm7.1 / deltanet none / conv none / narrow-gemv triton / stock-gemv triton / raw-gemv stock / decode eager", "platform": "rocm" }, "metrics": [ { "name": "compressed_decode_tok_s", "value": "10.63", "unit": "tok/s", "samples": [ "10.63", "10.61", "10.63" ] }, { "name": "compressed_prefill_tok_s", "value": "1145.02", "unit": "tok/s", "samples": [ "1144.18", "1153.37", "1145.02" ] }, { "name": "compressed_ttft_s", "value": "0.4673", "unit": "s", "samples": [ "0.4681", "0.4658", "0.4673" ] }, { "name": "read_gb_s", "value": "278.635", "unit": "GB/s" }, { "name": "copy_gb_s", "value": "250.01", "unit": "GB/s" }, { "name": "compressed_weights_gb", "value": "11.608", "unit": "GB" }, { "name": "compressed_decode_read_gb", "value": "10.33", "unit": "GB" } ], "model": { "name": "Qwen3-8B", "hfRepo": "Qwen/Qwen3-8B", "revision": "b968826d9c46dd6066d109eabc6255188de91218" }, "compression": { "profile": "gulp", "bitsPerWeight": "10.903", "meanTensorBitsPerWeight": "10.88", "packedBytes": 10314829151, "residentBytes": 11575275304 }, "stock": { "outcome": "skipped_predicted_nonfit", "budgetBytes": 17163091968 }, "site": { "source": "measurements/qwen3-8b_radeon-rx-7600-xt_2026-10-02_gulp.json", "at": { "compressed_decode_tok_s": "/metrics/0/value", "compressed_prefill_tok_s": "/metrics/1/value", "compressed_ttft_s": "/metrics/2/value", "read_gb_s": "/metrics/3/value", "copy_gb_s": "/metrics/4/value", "compressed_weights_gb": "/metrics/5/value", "compressed_decode_read_gb": "/metrics/6/value", "bitsPerWeight": "/compression/bitsPerWeight", "meanTensorBitsPerWeight": "/compression/meanTensorBitsPerWeight", "stock.outcome": "/stock/outcome", "stock.budgetBytes": "/stock/budgetBytes", "stockError": "/raw/stock_error", "allocatableGB": "/environment/memoryBytes" }, "allocatableGB": 17.163091968, "harness": "drinkme bench", "stockError": "skipped by bench's stock-arm fit check (predicted OOM under the stream loader's transient)" } } ] }