quantization_benchmarks.csv

model,quant method,truthfulqa_mc2 acc ± stderr,arc:challenge acc ± stderr,hellaswag acc,winogrande acc,average,throughput (tok/s),peak process vram (GB),calibration/quantization time,throughput w/ torch.compile,peak process vram w/ torch.compile,througput w/ marlin,peak process vram w/ marlin
Llama 3.1 8B,baseline (bf16),0.5456 ± 0.0150,0.5623 ± 0.0145,0.5973 ± 0.0049,0.7372 ± 0.0124,0.6106,38.82,16.77302,,79.27,17.12954,,
Llama 3.1 8B,AWQ (4 bit),0.5281 ± 0.0150,0.5589 ± .0145,.5869 ± .0049,0.7285 ± 0.0125,0.6006,43.09,8.11808,~10 minutes,,,,
Llama 3.1 8B,GPTQModel (4 bit),,,,,,36.83,6.31872,~23 minutes,,,37.84,6318.72
Llama 3.1 8B,AutoGPTQ (4 bit),0.5361 ± 0.0149,0.5589 ± 0.0145,0.5758 ± 0.0049,0.7103 ± 0.0127,0.595275,43.71,6.49278,~30 minutes,,,,
Llama 3.1 8B,bnb (nf4),0.5446 ± 0.0149,0.5563 ± 0.0145,0.5769 ± 0.0049,0.7301 ± 0.0125,0.601975,24.35,6.44035,~1 minute,,,,
Llama 3.1 8B,optimum quanto (w4a16),0.5162 ± 0.0150,0.5427 ± 0.0146,0.5856 ± 0.0049,0.7419 ± 0.0123,0.5966,31.22,6.57667,~30 seconds,,,,
Llama 3.1 8B,torchao (int4wo),0.5166 ±0.0150,0.5418 ± 0.0146,0.5855 ± 0.0049,0.7395 ± 0.0123,0.59585,24.98,6.49907,~20 seconds,85.76,6.8493,,
Llama 3.1 8B,HQQ (4 bit),0.5485 ± 0.0150,0.5495 ± 0.0145,0.5859 ± 0.0049,0.7285 ± 0.0125,0.6031,34.44,6.71718,,,,,
Llama 3.1 8B,HIGGS (4 bit),0.5214 ± 0.0149,0.5486 ± 0.0145,0.5836 ± 0.0049,0.7182 ± 0.0126,0.59295,28.35,6.81994,~5 minutes,,,,
Llama 3.1 8B,bnb (llm.int8()),0.5446 ± 0.0150,0.5623 ± 0.0145,0.5949 ± 0.0049,0.7324 ± 0.0124,0.60855,20.75,9.70772,~20 seconds,,,,
Llama 3.1 8B,HQQ (8 bit),0.5441 ± 0.0150, 0.5666 ± 0.0145, 0.5979 ± 0.0049, 0.7380 ± 0.0124,0.61165,9.07,10.6011,~80 seconds,,,,
Llama 3.1 8B,optimum quanto (int8wo),0.5436 ± 0.0150, 0.5640 ± 0.0145,0.5992 ± 0.0049, 0.7372 ± 0.0124,0.611,15.59,9.81887,~20 seconds,16.01,10.07052,,
Llama 3.1 8B,torchao (int8wo),0.5449 ± 0.0150,0.5640 ± 0.0145,0.5975 ± 0.0049,0.7380 ± 0.0124,0.6111,5.98,13.07155,~30 seconds,43.79,13.66714,,
Llama 3.1 8B,fbgemm (fp8),0.5430 ± 0.0150,0.5580 ± 0.0145,0.5958 ± 0.0049,0.7411 ± 0.0123,0.609475,33.83,10.00551,~30 seconds,,,,
Llama 3.1 8B,compressed-tensors (fp8),0.5398 ± 0.0151,0.5589 ± 0.0145,0.5950 ± 0.0049,0.7356 ± 0.0124,0.607325,,,,,,,
Llama 3.1 8B,VPTQ (2 bit),0.4543 ± 0.0149, 0.4923 ± 0.0146, 0.5258 ± 0.0050,0.6930 ± 0.0130,0.54135,32.35,5.28902,~2 hours,31.48,5.28692,,
Llama 3.1 8B,AQLM + PV (2 bit),0.5036 ± 0.0148,0.5230 ± 0.0146,0.5628 ± 0.0050,0.6938 ± 0.0130,0.5708,22.28,4.84023,~1 day,27.27,4.85491,,
Llama 3.1 8B,GPTQModel (2 bit),,,,,,19.02,18.45284,~26 minutes,,,,
Llama 3.1 8B,AutoGPTQ (2 bit),0.5127 ± 0.0150,0.1988 ± 0.0117,0.2665 ± 0.0044,0.4799 ± 0.0140,0.364475,6.25,11.02473,~26 minutes,,,,
Llama 3.1 70B,baseline (bf16),0.6068 ± 0.0147,0.6732 ± 0.0137,0.6666 ± 0.0047,0.8248 ± 0.0107,0.69285,9.73,142.26869,,10.1,142.81395,,
Llama 3.1 70B,AWQ (4 bit),0.5706 ± 0.0150,0.6681 ± 0.0138,0.6598 ± 0.0047,0.8193 ± 0.0108,0.67945,15.74,43.75288,~1 hour,,,,
Llama 3.1 70B,GPTQModel (4 bit),,,,,,14.84,40.5757,,,,15.28,40.5757
Llama 3.1 70B,AutoGPTQ (4 bit),0.5937 ± 0.0147,0.6655 ± 0.0138,0.6568 ± 0.0047,0.8185 ± 0.0108,0.683625,0.46,42.40022,~2 hours,,,,
Llama 3.1 70B,bnb (nf4),0.5939 ± 0.0148,0.6724 ± 0.0137,0.6592 ± 0.0047,0.8098 ± 0.0110,0.683825,11.27,44.62949,~2 minutes,,,,
Llama 3.1 70B,optimum quanto (w4a16),0.4847 ± 0.0164,0.2082 ± 0.0119,0.2582 ± 0.0044,0.4878 ± 0.0140,0.359725,12.97,80.39013,~2 minutes,,,,
Llama 3.1 70B,torchao (int4wo),0.4847 ± 0.0164,0.2108 ± 0.0119, 0.2581 ± 0.0044,  0.4980 ± 0.0141,0.3629,10.56,41.6054,~2 minutes,18.95,42.26181,,
Llama 3.1 70B,HQQ (4 bit),0.5882 ± 0.0146,0.6706 ± 0.0137, 0.6597 ± 0.0047,0.8035 ± 0.0112,0.6805,13.92,44.50366,~10 minutes,,,,
Llama 3.1 70B,HIGGS (4 bit),0.4871 ± 0.0163,0.1971 ± 0.0116,0.2575 ± 0.0044,0.4893 ± 0.0140,0.35775,11.61,41.52571,~6 minutes,12.38,41.02868,,
Llama 3.1 70B,bnb (llm.int8()),0.5604 ± 0.0169,0.6544 ± 0.0139,0.6382 ± 0.0048,0.7940 ± 0.0114,0.66175,6.87,74.26428,~2 minutes,,,,
Llama 3.1 70B,HQQ (8 bit),0.6112 ± 0.0146,0.6732 ± 0.0137,0.6661 ± 0.0047,0.8327 ± 0.0105,0.6958,0.98,80.52435,~10 minutes,0.98,80.39013,,
Llama 3.1 70B,optimum quanto (int8wo),0.5591 ± 0.0150,0.6459 ± 0.0140,0.6413 ± 0.0048,0.7979 ± 0.0113,0.66105,1.79,74.21192,~2 minutes,1.8,74.21401,,
Llama 3.1 70B,torchao (int8wo),0.6094 ± 0.0146,0.6732 ± 0.0137,0.6659 ± 0.0047,0.8240 ± 0.0107,0.693125,0.65,89.85038,~2 minutes,0.65,89.84619,,
Llama 3.1 70B,fbgemm (fp8),0.6075 ± 0.0146,0.6732 ± 0.0137,0.6671 ± 0.0047,0.8216 ± 0.0108,0.69235,13.61,74.04624,~6 minutes,,,,
Llama 3.1 70B,compressed-tensors (fp8),0.6062 ± 0.0146,0.6741 ± 0.0137,0.6652 ± 0.0047,0.8216 ± 0.0108,0.691775,,,,,,,
Llama 3.1 70B,VPTQ (2 bit),0.5451 ± 0.0150, 0.6212 ± 0.0142,0.6073 ± 0.0049, 0.7901 ± 0.0114,0.640925,6.29,24.89949,~19 hours,6.18,24.89949,,
Llama 3.1 70B,AQLM + PV (2 bit),0.5706 ± 0.0150,0.6365 ± 0.0141,0.6401 ± 0.0048,0.8066 ± 0.0111,0.66345,6.75,23.12739,10-14 days,7.09,23607.64,,
Llama 3.1 70B,GPTQModel (2 bit),,,,,,,,,,,,
Llama 3.1 70B,AutoGPTQ (2 bit),0.4556 ± 0.0147,0.2807 ± 0.0131,0.3642 ± 0.0048,0.5470 ± 0.0140,0.411875,,,,,,,