{ "gptq": [ { "process": "gptq", "layer": 0, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000640668", "samples": "46025", "damp": "0.01000", "time": "4.484", "fwd_time": "1.254", "(v)ram": "cuda 3.84G" }, { "process": "gptq", "layer": 0, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000425612", "samples": "46025", "damp": "0.01000", "time": "4.580", "fwd_time": "0.569", "(v)ram": "cuda 3.99G" }, { "process": "gptq", "layer": 0, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000992", "samples": "46025", "damp": "0.01000", "time": "9.272", "fwd_time": "1.231", "(v)ram": "cuda 4.15G" }, { "process": "gptq", "layer": 0, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000004416", "samples": "46025", "damp": "0.01000", "time": "11.330", "fwd_time": "1.887", "(v)ram": "cuda 6.49G" }, { "process": "gptq", "layer": 0, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000004125", "samples": "46025", "damp": "0.01000", "time": "11.633", "fwd_time": "1.887", "(v)ram": "cuda 6.49G" }, { "process": "gptq", "layer": 0, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000015", "samples": "46025", "damp": "0.01000", "time": "10.874", "fwd_time": "2.786", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 1, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000066808", "samples": "46025", "damp": "0.01000", "time": "5.618", "fwd_time": "0.591", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 1, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000041003", "samples": "46025", "damp": "0.01000", "time": "5.369", "fwd_time": "0.567", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 1, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000099", "samples": "46025", "damp": "0.01000", "time": "8.345", "fwd_time": "1.188", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 1, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000008680", "samples": "46025", "damp": "0.01000", "time": "11.915", "fwd_time": "1.712", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 1, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000009028", "samples": "46025", "damp": "0.01000", "time": "12.077", "fwd_time": "1.712", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 1, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000023", "samples": "46025", "damp": "0.01000", "time": "17.961", "fwd_time": "2.866", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 2, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000090825", "samples": "46025", "damp": "0.01000", "time": "4.822", "fwd_time": "0.841", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 2, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000055125", "samples": "46025", "damp": "0.01000", "time": "-2.177", "fwd_time": "0.582", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 2, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000157", "samples": "46025", "damp": "0.01000", "time": "5.873", "fwd_time": "1.405", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 2, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000017131", "samples": "46025", "damp": "0.01000", "time": "12.168", "fwd_time": "1.688", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 2, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000016172", "samples": "46025", "damp": "0.01000", "time": "12.324", "fwd_time": "1.688", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 2, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000074", "samples": "46025", "damp": "0.01000", "time": "16.749", "fwd_time": "2.946", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 3, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000096932", "samples": "46025", "damp": "0.01000", "time": "18.543", "fwd_time": "0.891", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 3, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0001398233", "samples": "46025", "damp": "0.01000", "time": "18.691", "fwd_time": "0.891", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 3, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000094358", "samples": "46025", "damp": "0.01000", "time": "27.247", "fwd_time": "0.891", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 3, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000209", "samples": "46025", "damp": "0.01000", "time": "4.958", "fwd_time": "0.954", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 3, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000019242", "samples": "46025", "damp": "0.01000", "time": "10.999", "fwd_time": "1.443", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 3, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000019810", "samples": "46025", "damp": "0.01000", "time": "11.184", "fwd_time": "1.443", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 3, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000055", "samples": "46025", "damp": "0.01000", "time": "12.275", "fwd_time": "2.669", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 4, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000094861", "samples": "46025", "damp": "0.01000", "time": "4.077", "fwd_time": "0.598", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 4, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000060512", "samples": "46025", "damp": "0.01000", "time": "3.983", "fwd_time": "0.556", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 4, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000215", "samples": "46025", "damp": "0.01000", "time": "4.609", "fwd_time": "1.202", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 4, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000026429", "samples": "46025", "damp": "0.01000", "time": "11.536", "fwd_time": "1.635", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 4, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000026010", "samples": "46025", "damp": "0.01000", "time": "11.805", "fwd_time": "1.635", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 4, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000073", "samples": "46025", "damp": "0.01000", "time": "16.352", "fwd_time": "3.006", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 5, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000133733", "samples": "46025", "damp": "0.01000", "time": "3.602", "fwd_time": "0.581", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 5, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000087068", "samples": "46025", "damp": "0.01000", "time": "3.274", "fwd_time": "0.496", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 5, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000237", "samples": "46025", "damp": "0.01000", "time": "4.176", "fwd_time": "1.168", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 5, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000032260", "samples": "46025", "damp": "0.01000", "time": "9.491", "fwd_time": "1.588", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 5, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000033839", "samples": "46025", "damp": "0.01000", "time": "9.740", "fwd_time": "1.588", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 5, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000106", "samples": "46025", "damp": "0.01000", "time": "15.043", "fwd_time": "3.020", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 6, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000238767", "samples": "46025", "damp": "0.01000", "time": "2.756", "fwd_time": "0.561", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 6, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000142515", "samples": "46025", "damp": "0.01000", "time": "3.279", "fwd_time": "0.481", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 6, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000460", "samples": "46025", "damp": "0.01000", "time": "4.376", "fwd_time": "1.119", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 6, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000050037", "samples": "46025", "damp": "0.01000", "time": "18.529", "fwd_time": "1.570", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 6, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000047293", "samples": "46025", "damp": "0.01000", "time": "19.026", "fwd_time": "1.570", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 6, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000190", "samples": "46025", "damp": "0.01000", "time": "13.388", "fwd_time": "3.030", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 7, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000081105", "samples": "46025", "damp": "0.01000", "time": "7.551", "fwd_time": "1.150", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 7, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000077117", "samples": "46025", "damp": "0.01000", "time": "7.710", "fwd_time": "1.150", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 7, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0001078202", "samples": "46025", "damp": "0.01000", "time": "7.939", "fwd_time": "1.150", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 7, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000533", "samples": "46025", "damp": "0.01000", "time": "4.207", "fwd_time": "0.822", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 7, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000052691", "samples": "46025", "damp": "0.01000", "time": "21.493", "fwd_time": "1.251", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 7, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000055773", "samples": "46025", "damp": "0.01000", "time": "21.837", "fwd_time": "1.251", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 7, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000240", "samples": "46025", "damp": "0.01000", "time": "19.025", "fwd_time": "2.856", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 8, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000194353", "samples": "46025", "damp": "0.01000", "time": "3.974", "fwd_time": "0.646", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 8, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000122614", "samples": "46025", "damp": "0.01000", "time": "3.839", "fwd_time": "0.488", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 8, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000606", "samples": "46025", "damp": "0.01000", "time": "4.391", "fwd_time": "1.104", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 8, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000058376", "samples": "46025", "damp": "0.01000", "time": "12.392", "fwd_time": "1.479", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 8, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000061431", "samples": "46025", "damp": "0.01000", "time": "12.472", "fwd_time": "1.479", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 8, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000260", "samples": "46025", "damp": "0.01000", "time": "20.321", "fwd_time": "3.119", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 9, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000203141", "samples": "46025", "damp": "0.01000", "time": "3.038", "fwd_time": "0.552", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 9, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000124319", "samples": "46025", "damp": "0.01000", "time": "3.160", "fwd_time": "0.473", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 9, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000581", "samples": "46025", "damp": "0.01000", "time": "4.044", "fwd_time": "1.152", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 9, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000063746", "samples": "46025", "damp": "0.01000", "time": "10.923", "fwd_time": "1.610", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 9, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000060999", "samples": "46025", "damp": "0.01000", "time": "11.014", "fwd_time": "1.610", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 9, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000285", "samples": "46025", "damp": "0.01000", "time": "17.955", "fwd_time": "3.109", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 10, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000206230", "samples": "46025", "damp": "0.01000", "time": "3.743", "fwd_time": "0.656", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 10, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000123166", "samples": "46025", "damp": "0.01000", "time": "3.892", "fwd_time": "0.548", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 10, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000868", "samples": "46025", "damp": "0.01000", "time": "-4.116", "fwd_time": "1.141", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 10, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000064383", "samples": "46025", "damp": "0.01000", "time": "9.809", "fwd_time": "1.594", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 10, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000062146", "samples": "46025", "damp": "0.01000", "time": "9.850", "fwd_time": "1.594", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 10, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000342", "samples": "46025", "damp": "0.01000", "time": "16.094", "fwd_time": "3.075", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 11, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000065789", "samples": "46025", "damp": "0.01000", "time": "16.954", "fwd_time": "0.995", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 11, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000068888", "samples": "46025", "damp": "0.01000", "time": "25.298", "fwd_time": "0.995", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 11, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0000868339", "samples": "46025", "damp": "0.01000", "time": "17.150", "fwd_time": "0.995", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 11, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000001428", "samples": "46025", "damp": "0.01000", "time": "3.949", "fwd_time": "0.905", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 11, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000074936", "samples": "46025", "damp": "0.01000", "time": "9.433", "fwd_time": "1.326", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 11, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000072303", "samples": "46025", "damp": "0.01000", "time": "9.511", "fwd_time": "1.326", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 11, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000487", "samples": "46025", "damp": "0.01000", "time": "16.934", "fwd_time": "2.786", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 12, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000268114", "samples": "46025", "damp": "0.01000", "time": "3.592", "fwd_time": "0.637", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 12, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000159200", "samples": "46025", "damp": "0.01000", "time": "3.598", "fwd_time": "0.516", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 12, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000001276", "samples": "46025", "damp": "0.01000", "time": "-5.044", "fwd_time": "1.149", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 12, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000075678", "samples": "46025", "damp": "0.01000", "time": "9.779", "fwd_time": "1.677", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 12, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000078482", "samples": "46025", "damp": "0.01000", "time": "9.936", "fwd_time": "1.677", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 12, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000544", "samples": "46025", "damp": "0.01000", "time": "15.075", "fwd_time": "3.041", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 13, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000284073", "samples": "46025", "damp": "0.01000", "time": "3.514", "fwd_time": "0.662", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 13, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000160394", "samples": "46025", "damp": "0.01000", "time": "3.355", "fwd_time": "0.508", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 13, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000001497", "samples": "46025", "damp": "0.01000", "time": "4.377", "fwd_time": "1.163", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 13, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000086531", "samples": "46025", "damp": "0.01000", "time": "18.102", "fwd_time": "1.636", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 13, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000083314", "samples": "46025", "damp": "0.01000", "time": "18.183", "fwd_time": "1.636", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 13, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000629", "samples": "46025", "damp": "0.01000", "time": "4.334", "fwd_time": "2.996", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 14, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000316271", "samples": "46025", "damp": "0.01000", "time": "3.108", "fwd_time": "0.645", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 14, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000186084", "samples": "46025", "damp": "0.01000", "time": "11.488", "fwd_time": "0.527", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 14, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000002475", "samples": "46025", "damp": "0.01000", "time": "4.033", "fwd_time": "1.114", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 14, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000091396", "samples": "46025", "damp": "0.01000", "time": "9.448", "fwd_time": "1.482", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 14, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000094544", "samples": "46025", "damp": "0.01000", "time": "9.787", "fwd_time": "1.482", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 14, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000797", "samples": "46025", "damp": "0.01000", "time": "13.106", "fwd_time": "2.798", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 15, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000058696", "samples": "46025", "damp": "0.01000", "time": "26.247", "fwd_time": "0.892", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 15, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0000685723", "samples": "46025", "damp": "0.01000", "time": "26.544", "fwd_time": "0.892", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 15, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000059185", "samples": "46025", "damp": "0.01000", "time": "26.619", "fwd_time": "0.892", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 15, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000001955", "samples": "46025", "damp": "0.01000", "time": "3.753", "fwd_time": "1.037", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 15, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000113858", "samples": "46025", "damp": "0.01000", "time": "9.760", "fwd_time": "1.185", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 15, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000111234", "samples": "46025", "damp": "0.01000", "time": "9.954", "fwd_time": "1.185", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 15, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000001126", "samples": "46025", "damp": "0.01000", "time": "14.539", "fwd_time": "2.617", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 16, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000340021", "samples": "46025", "damp": "0.01000", "time": "3.784", "fwd_time": "0.544", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 16, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000193314", "samples": "46025", "damp": "0.01000", "time": "4.452", "fwd_time": "0.523", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 16, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000002341", "samples": "46025", "damp": "0.01000", "time": "5.521", "fwd_time": "1.128", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 16, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000125576", "samples": "46025", "damp": "0.01000", "time": "11.582", "fwd_time": "1.567", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 16, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000129008", "samples": "46025", "damp": "0.01000", "time": "12.202", "fwd_time": "1.567", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 16, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000001454", "samples": "46025", "damp": "0.01000", "time": "17.559", "fwd_time": "2.978", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 17, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000392903", "samples": "46025", "damp": "0.01000", "time": "-4.769", "fwd_time": "0.526", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 17, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000201683", "samples": "46025", "damp": "0.01000", "time": "4.605", "fwd_time": "0.540", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 17, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000003138", "samples": "46025", "damp": "0.01000", "time": "13.516", "fwd_time": "1.085", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 17, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000153182", "samples": "46025", "damp": "0.01000", "time": "12.437", "fwd_time": "1.661", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 17, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000149272", "samples": "46025", "damp": "0.01000", "time": "12.847", "fwd_time": "1.661", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 17, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000001908", "samples": "46025", "damp": "0.01000", "time": "18.662", "fwd_time": "2.865", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 18, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000423103", "samples": "46025", "damp": "0.01000", "time": "2.916", "fwd_time": "0.497", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 18, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000235245", "samples": "46025", "damp": "0.01000", "time": "-4.889", "fwd_time": "0.463", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 18, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000005078", "samples": "46025", "damp": "0.01000", "time": "4.151", "fwd_time": "1.101", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 18, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000193033", "samples": "46025", "damp": "0.01000", "time": "9.565", "fwd_time": "1.469", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 18, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000198968", "samples": "46025", "damp": "0.01000", "time": "9.763", "fwd_time": "1.469", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 18, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000003688", "samples": "46025", "damp": "0.01000", "time": "17.496", "fwd_time": "2.724", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 19, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000068886", "samples": "46025", "damp": "0.01000", "time": "17.284", "fwd_time": "0.897", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 19, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000063744", "samples": "46025", "damp": "0.01000", "time": "17.370", "fwd_time": "0.897", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 19, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0000750437", "samples": "46025", "damp": "0.01000", "time": "17.558", "fwd_time": "0.897", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 19, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000012625", "samples": "46025", "damp": "0.01000", "time": "4.216", "fwd_time": "0.908", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 19, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000293955", "samples": "46025", "damp": "0.01000", "time": "9.437", "fwd_time": "1.245", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 19, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000303575", "samples": "46025", "damp": "0.01000", "time": "9.640", "fwd_time": "1.245", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 19, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000006327", "samples": "46025", "damp": "0.01000", "time": "17.143", "fwd_time": "2.498", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 20, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000871094", "samples": "46025", "damp": "0.01000", "time": "3.164", "fwd_time": "0.581", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 20, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000497236", "samples": "46025", "damp": "0.01000", "time": "3.134", "fwd_time": "0.467", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 20, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000011158", "samples": "46025", "damp": "0.01000", "time": "-4.141", "fwd_time": "1.051", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 20, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000392667", "samples": "46025", "damp": "0.01000", "time": "11.414", "fwd_time": "1.630", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 20, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000372829", "samples": "46025", "damp": "0.01000", "time": "11.600", "fwd_time": "1.630", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 20, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000008790", "samples": "46025", "damp": "0.01000", "time": "26.521", "fwd_time": "3.115", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 21, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0001053623", "samples": "46025", "damp": "0.01000", "time": "3.435", "fwd_time": "0.605", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 21, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000681308", "samples": "46025", "damp": "0.01000", "time": "3.351", "fwd_time": "0.494", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 21, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000012490", "samples": "46025", "damp": "0.01000", "time": "3.787", "fwd_time": "1.113", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 21, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000423605", "samples": "46025", "damp": "0.01000", "time": "8.915", "fwd_time": "1.535", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 21, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000444869", "samples": "46025", "damp": "0.01000", "time": "9.126", "fwd_time": "1.535", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 21, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000008031", "samples": "46025", "damp": "0.01000", "time": "13.210", "fwd_time": "2.559", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 22, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000931988", "samples": "46025", "damp": "0.01000", "time": "5.057", "fwd_time": "0.489", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 22, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000577565", "samples": "46025", "damp": "0.01000", "time": "5.946", "fwd_time": "0.443", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 22, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000009584", "samples": "46025", "damp": "0.01000", "time": "6.068", "fwd_time": "1.132", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 22, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000540310", "samples": "46025", "damp": "0.01000", "time": "10.393", "fwd_time": "1.615", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 22, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000489587", "samples": "46025", "damp": "0.01000", "time": "10.421", "fwd_time": "1.615", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 22, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000009408", "samples": "46025", "damp": "0.01000", "time": "13.586", "fwd_time": "2.861", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 23, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000109871", "samples": "46025", "damp": "0.01000", "time": "18.779", "fwd_time": "1.014", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 23, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0001100955", "samples": "46025", "damp": "0.01000", "time": "19.032", "fwd_time": "1.014", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 23, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000098417", "samples": "46025", "damp": "0.01000", "time": "19.288", "fwd_time": "1.014", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 23, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000016605", "samples": "46025", "damp": "0.01000", "time": "4.038", "fwd_time": "0.976", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 23, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000644726", "samples": "46025", "damp": "0.01000", "time": "9.764", "fwd_time": "1.330", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 23, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000607319", "samples": "46025", "damp": "0.01000", "time": "9.959", "fwd_time": "1.330", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 23, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000011916", "samples": "46025", "damp": "0.01000", "time": "12.735", "fwd_time": "2.692", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 24, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000977541", "samples": "46025", "damp": "0.01000", "time": "5.026", "fwd_time": "0.668", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 24, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000563157", "samples": "46025", "damp": "0.01000", "time": "4.291", "fwd_time": "0.566", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 24, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000017884", "samples": "46025", "damp": "0.01000", "time": "5.123", "fwd_time": "1.223", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 24, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000749009", "samples": "46025", "damp": "0.01000", "time": "9.060", "fwd_time": "1.721", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 24, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000704703", "samples": "46025", "damp": "0.01000", "time": "9.230", "fwd_time": "1.721", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 24, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000015496", "samples": "46025", "damp": "0.01000", "time": "12.893", "fwd_time": "2.971", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 25, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0001061138", "samples": "46025", "damp": "0.01000", "time": "3.294", "fwd_time": "0.564", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 25, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000575017", "samples": "46025", "damp": "0.01000", "time": "3.575", "fwd_time": "0.504", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 25, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000019334", "samples": "46025", "damp": "0.01000", "time": "-5.667", "fwd_time": "1.233", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 25, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000728622", "samples": "46025", "damp": "0.01000", "time": "11.275", "fwd_time": "1.625", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 25, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000718050", "samples": "46025", "damp": "0.01000", "time": "11.798", "fwd_time": "1.625", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 25, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000017222", "samples": "46025", "damp": "0.01000", "time": "12.265", "fwd_time": "3.080", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 26, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000980629", "samples": "46025", "damp": "0.01000", "time": "3.373", "fwd_time": "0.559", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 26, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000513086", "samples": "46025", "damp": "0.01000", "time": "3.949", "fwd_time": "0.573", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 26, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000031564", "samples": "46025", "damp": "0.01000", "time": "4.286", "fwd_time": "1.135", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 26, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000728342", "samples": "46025", "damp": "0.01000", "time": "10.045", "fwd_time": "1.568", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 26, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000751999", "samples": "46025", "damp": "0.01000", "time": "10.461", "fwd_time": "1.568", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 26, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000022494", "samples": "46025", "damp": "0.01000", "time": "14.526", "fwd_time": "3.086", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 27, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000114680", "samples": "46025", "damp": "0.01000", "time": "18.320", "fwd_time": "0.734", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 27, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000110706", "samples": "46025", "damp": "0.01000", "time": "18.448", "fwd_time": "0.734", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 27, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0001244631", "samples": "46025", "damp": "0.01000", "time": "18.561", "fwd_time": "0.734", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 27, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000056642", "samples": "46025", "damp": "0.01000", "time": "3.812", "fwd_time": "0.771", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 27, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000887116", "samples": "46025", "damp": "0.01000", "time": "11.084", "fwd_time": "1.233", "(v)ram": "cuda 11.02G" }, { "process": "gptq", "layer": 27, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000947490", "samples": "46025", "damp": "0.01000", "time": "11.226", "fwd_time": "1.233", "(v)ram": "cuda 11.02G" }, { "process": "gptq", "layer": 27, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000032511", "samples": "46025", "damp": "0.01000", "time": "16.286", "fwd_time": "2.815", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 28, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0001241501", "samples": "46025", "damp": "0.01000", "time": "3.169", "fwd_time": "0.571", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 28, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000596245", "samples": "46025", "damp": "0.01000", "time": "4.270", "fwd_time": "0.498", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 28, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000040309", "samples": "46025", "damp": "0.01000", "time": "4.836", "fwd_time": "1.146", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 28, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0001008322", "samples": "46025", "damp": "0.01000", "time": "10.028", "fwd_time": "1.747", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 28, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000927496", "samples": "46025", "damp": "0.01000", "time": "1.926", "fwd_time": "1.747", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 28, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000035256", "samples": "46025", "damp": "0.01000", "time": "18.627", "fwd_time": "3.047", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 29, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0001449247", "samples": "46025", "damp": "0.01000", "time": "3.526", "fwd_time": "0.597", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 29, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000673718", "samples": "46025", "damp": "0.01000", "time": "3.038", "fwd_time": "0.529", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 29, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000043401", "samples": "46025", "damp": "0.01000", "time": "4.864", "fwd_time": "1.210", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 29, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0001089125", "samples": "46025", "damp": "0.01000", "time": "10.248", "fwd_time": "1.702", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 29, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000997183", "samples": "46025", "damp": "0.01000", "time": "10.289", "fwd_time": "1.702", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 29, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000042763", "samples": "46025", "damp": "0.01000", "time": "18.978", "fwd_time": "3.102", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 30, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0001554490", "samples": "46025", "damp": "0.01000", "time": "3.582", "fwd_time": "0.581", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 30, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000741639", "samples": "46025", "damp": "0.01000", "time": "3.510", "fwd_time": "0.498", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 30, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000069002", "samples": "46025", "damp": "0.01000", "time": "4.910", "fwd_time": "1.209", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 30, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0001074855", "samples": "46025", "damp": "0.01000", "time": "11.820", "fwd_time": "1.822", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 30, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0001199021", "samples": "46025", "damp": "0.01000", "time": "12.061", "fwd_time": "1.822", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 30, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000058043", "samples": "46025", "damp": "0.01000", "time": "23.391", "fwd_time": "3.133", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 31, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000169266", "samples": "46025", "damp": "0.01000", "time": "17.693", "fwd_time": "1.032", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 31, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000209107", "samples": "46025", "damp": "0.01000", "time": "18.069", "fwd_time": "1.032", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 31, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0001738664", "samples": "46025", "damp": "0.01000", "time": "18.220", "fwd_time": "1.032", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 31, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000113816", "samples": "46025", "damp": "0.01000", "time": "4.668", "fwd_time": "1.015", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 31, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0001441583", "samples": "46025", "damp": "0.01000", "time": "10.405", "fwd_time": "1.410", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 31, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0001267508", "samples": "46025", "damp": "0.01000", "time": "10.597", "fwd_time": "1.410", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 31, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000077594", "samples": "46025", "damp": "0.01000", "time": "16.786", "fwd_time": "2.772", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 32, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0002006095", "samples": "46025", "damp": "0.01000", "time": "3.680", "fwd_time": "0.699", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 32, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000890080", "samples": "46025", "damp": "0.01000", "time": "3.283", "fwd_time": "0.511", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 32, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000086844", "samples": "46025", "damp": "0.01000", "time": "4.131", "fwd_time": "1.147", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 32, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0001529837", "samples": "46025", "damp": "0.01000", "time": "18.950", "fwd_time": "1.634", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 32, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0001750664", "samples": "46025", "damp": "0.01000", "time": "19.777", "fwd_time": "1.634", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 32, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000093934", "samples": "46025", "damp": "0.01000", "time": "5.067", "fwd_time": "3.016", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 33, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0002594677", "samples": "46025", "damp": "0.01000", "time": "3.418", "fwd_time": "0.628", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 33, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0001079468", "samples": "46025", "damp": "0.01000", "time": "3.559", "fwd_time": "0.473", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 33, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000115782", "samples": "46025", "damp": "0.01000", "time": "4.137", "fwd_time": "1.176", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 33, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0001775377", "samples": "46025", "damp": "0.01000", "time": "10.220", "fwd_time": "1.596", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 33, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002057817", "samples": "46025", "damp": "0.01000", "time": "10.360", "fwd_time": "1.596", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 33, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000103116", "samples": "46025", "damp": "0.01000", "time": "15.734", "fwd_time": "3.053", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 34, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0002748349", "samples": "46025", "damp": "0.01000", "time": "4.060", "fwd_time": "0.735", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 34, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0001253682", "samples": "46025", "damp": "0.01000", "time": "4.193", "fwd_time": "0.495", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 34, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000114199", "samples": "46025", "damp": "0.01000", "time": "4.533", "fwd_time": "1.200", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 34, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0001920355", "samples": "46025", "damp": "0.01000", "time": "17.583", "fwd_time": "1.581", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 34, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002284183", "samples": "46025", "damp": "0.01000", "time": "9.413", "fwd_time": "1.581", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 34, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000137695", "samples": "46025", "damp": "0.01000", "time": "16.264", "fwd_time": "3.059", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 35, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000287168", "samples": "46025", "damp": "0.01000", "time": "19.131", "fwd_time": "0.889", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 35, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000441700", "samples": "46025", "damp": "0.01000", "time": "19.177", "fwd_time": "0.889", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 35, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0003003337", "samples": "46025", "damp": "0.01000", "time": "19.536", "fwd_time": "0.889", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 35, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000242998", "samples": "46025", "damp": "0.01000", "time": "3.689", "fwd_time": "1.025", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 35, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002463556", "samples": "46025", "damp": "0.01000", "time": "10.482", "fwd_time": "1.340", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 35, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002163312", "samples": "46025", "damp": "0.01000", "time": "10.775", "fwd_time": "1.340", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 35, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000127315", "samples": "46025", "damp": "0.01000", "time": "18.150", "fwd_time": "2.846", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 36, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0003812556", "samples": "46025", "damp": "0.01000", "time": "3.222", "fwd_time": "0.623", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 36, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0001943348", "samples": "46025", "damp": "0.01000", "time": "3.575", "fwd_time": "0.527", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 36, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000062425", "samples": "46025", "damp": "0.01000", "time": "3.888", "fwd_time": "1.209", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 36, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002391412", "samples": "46025", "damp": "0.01000", "time": "9.218", "fwd_time": "1.624", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 36, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002363071", "samples": "46025", "damp": "0.01000", "time": "9.388", "fwd_time": "1.624", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 36, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000098985", "samples": "46025", "damp": "0.01000", "time": "14.704", "fwd_time": "3.084", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 37, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0003956966", "samples": "46025", "damp": "0.01000", "time": "3.865", "fwd_time": "0.675", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 37, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0002285038", "samples": "46025", "damp": "0.01000", "time": "12.878", "fwd_time": "0.568", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 37, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000066447", "samples": "46025", "damp": "0.01000", "time": "5.416", "fwd_time": "1.216", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 37, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002234319", "samples": "46025", "damp": "0.01000", "time": "0.567", "fwd_time": "1.702", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 37, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002264691", "samples": "46025", "damp": "0.01000", "time": "0.646", "fwd_time": "1.702", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 37, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000087398", "samples": "46025", "damp": "0.01000", "time": "13.288", "fwd_time": "3.051", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 38, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0003584955", "samples": "46025", "damp": "0.01000", "time": "3.428", "fwd_time": "0.666", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 38, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0002117160", "samples": "46025", "damp": "0.01000", "time": "3.429", "fwd_time": "0.485", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 38, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000049733", "samples": "46025", "damp": "0.01000", "time": "5.615", "fwd_time": "1.209", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 38, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002952798", "samples": "46025", "damp": "0.01000", "time": "12.555", "fwd_time": "1.811", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 38, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002543049", "samples": "46025", "damp": "0.01000", "time": "12.622", "fwd_time": "1.811", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 38, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000092283", "samples": "46025", "damp": "0.01000", "time": "13.004", "fwd_time": "3.051", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 39, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000371762", "samples": "46025", "damp": "0.01000", "time": "15.638", "fwd_time": "0.994", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 39, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000277842", "samples": "46025", "damp": "0.01000", "time": "16.113", "fwd_time": "0.994", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 39, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0003053927", "samples": "46025", "damp": "0.01000", "time": "16.328", "fwd_time": "0.994", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 39, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000147484", "samples": "46025", "damp": "0.01000", "time": "5.418", "fwd_time": "1.025", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 39, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002592258", "samples": "46025", "damp": "0.01000", "time": "19.799", "fwd_time": "1.473", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 39, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002373144", "samples": "46025", "damp": "0.01000", "time": "11.591", "fwd_time": "1.473", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 39, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000076080", "samples": "46025", "damp": "0.01000", "time": "11.489", "fwd_time": "2.849", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 40, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0003056463", "samples": "46025", "damp": "0.01000", "time": "3.870", "fwd_time": "0.608", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 40, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0001710136", "samples": "46025", "damp": "0.01000", "time": "3.869", "fwd_time": "0.548", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 40, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000064311", "samples": "46025", "damp": "0.01000", "time": "3.865", "fwd_time": "1.139", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 40, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002815306", "samples": "46025", "damp": "0.01000", "time": "9.036", "fwd_time": "1.622", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 40, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002530349", "samples": "46025", "damp": "0.01000", "time": "9.325", "fwd_time": "1.622", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 40, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000087425", "samples": "46025", "damp": "0.01000", "time": "17.604", "fwd_time": "2.997", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 41, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0002874149", "samples": "46025", "damp": "0.01000", "time": "2.936", "fwd_time": "0.643", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 41, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0001601348", "samples": "46025", "damp": "0.01000", "time": "3.490", "fwd_time": "0.506", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 41, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000050767", "samples": "46025", "damp": "0.01000", "time": "4.408", "fwd_time": "1.124", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 41, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002341963", "samples": "46025", "damp": "0.01000", "time": "9.591", "fwd_time": "1.673", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 41, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002359249", "samples": "46025", "damp": "0.01000", "time": "9.602", "fwd_time": "1.673", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 41, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000094304", "samples": "46025", "damp": "0.01000", "time": "13.578", "fwd_time": "3.016", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 42, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0002636356", "samples": "46025", "damp": "0.01000", "time": "3.326", "fwd_time": "0.581", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 42, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0001423076", "samples": "46025", "damp": "0.01000", "time": "3.340", "fwd_time": "0.504", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 42, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000093059", "samples": "46025", "damp": "0.01000", "time": "3.949", "fwd_time": "1.131", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 42, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002160006", "samples": "46025", "damp": "0.01000", "time": "7.731", "fwd_time": "1.618", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 42, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002293094", "samples": "46025", "damp": "0.01000", "time": "7.876", "fwd_time": "1.618", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 42, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000119015", "samples": "46025", "damp": "0.01000", "time": "13.831", "fwd_time": "3.092", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 43, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000283694", "samples": "46025", "damp": "0.01000", "time": "18.851", "fwd_time": "0.958", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 43, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000386960", "samples": "46025", "damp": "0.01000", "time": "19.225", "fwd_time": "0.958", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 43, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0002999400", "samples": "46025", "damp": "0.01000", "time": "19.685", "fwd_time": "0.958", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 43, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000282183", "samples": "46025", "damp": "0.01000", "time": "4.394", "fwd_time": "1.054", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 43, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002412899", "samples": "46025", "damp": "0.01000", "time": "8.769", "fwd_time": "1.336", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 43, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002669067", "samples": "46025", "damp": "0.01000", "time": "8.915", "fwd_time": "1.336", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 43, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000145095", "samples": "46025", "damp": "0.01000", "time": "14.120", "fwd_time": "2.744", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 44, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0003010677", "samples": "46025", "damp": "0.01000", "time": "3.444", "fwd_time": "0.545", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 44, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0001483966", "samples": "46025", "damp": "0.01000", "time": "3.424", "fwd_time": "0.476", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 44, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000109336", "samples": "46025", "damp": "0.01000", "time": "6.900", "fwd_time": "1.209", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 44, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002777499", "samples": "46025", "damp": "0.01000", "time": "10.031", "fwd_time": "1.527", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 44, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002487458", "samples": "46025", "damp": "0.01000", "time": "10.118", "fwd_time": "1.527", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 44, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000153657", "samples": "46025", "damp": "0.01000", "time": "14.335", "fwd_time": "2.905", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 45, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0003254824", "samples": "46025", "damp": "0.01000", "time": "3.175", "fwd_time": "0.497", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 45, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0001588313", "samples": "46025", "damp": "0.01000", "time": "3.163", "fwd_time": "0.460", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 45, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000099234", "samples": "46025", "damp": "0.01000", "time": "3.790", "fwd_time": "1.114", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 45, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002550306", "samples": "46025", "damp": "0.01000", "time": "12.429", "fwd_time": "1.659", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 45, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002858416", "samples": "46025", "damp": "0.01000", "time": "12.793", "fwd_time": "1.659", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 45, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000168444", "samples": "46025", "damp": "0.01000", "time": "13.149", "fwd_time": "2.835", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 46, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0003418054", "samples": "46025", "damp": "0.01000", "time": "3.455", "fwd_time": "0.647", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 46, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0001685697", "samples": "46025", "damp": "0.01000", "time": "3.348", "fwd_time": "0.517", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 46, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000159044", "samples": "46025", "damp": "0.01000", "time": "3.975", "fwd_time": "1.098", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 46, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002585288", "samples": "46025", "damp": "0.01000", "time": "11.718", "fwd_time": "1.514", "(v)ram": "cuda 11.04G" }, { "process": "gptq", "layer": 46, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0002961667", "samples": "46025", "damp": "0.01000", "time": "12.021", "fwd_time": "1.514", "(v)ram": "cuda 11.04G" }, { "process": "gptq", "layer": 46, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000194553", "samples": "46025", "damp": "0.01000", "time": "21.164", "fwd_time": "3.161", "(v)ram": "cuda 11.03G" }, { "process": "gptq", "layer": 47, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000661487", "samples": "46025", "damp": "0.01000", "time": "15.668", "fwd_time": "0.958", "(v)ram": "cuda 11.02G" }, { "process": "gptq", "layer": 47, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0003986497", "samples": "46025", "damp": "0.01000", "time": "16.095", "fwd_time": "0.958", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 47, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000383166", "samples": "46025", "damp": "0.01000", "time": "16.362", "fwd_time": "0.958", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 47, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000384544", "samples": "46025", "damp": "0.01000", "time": "5.072", "fwd_time": "0.834", "(v)ram": "cuda 11.02G" }, { "process": "gptq", "layer": 47, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0003089703", "samples": "46025", "damp": "0.01000", "time": "11.450", "fwd_time": "1.242", "(v)ram": "cuda 11.03G" }, { "process": "gptq", "layer": 47, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0003527128", "samples": "46025", "damp": "0.01000", "time": "11.739", "fwd_time": "1.242", "(v)ram": "cuda 11.03G" }, { "process": "gptq", "layer": 47, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000250715", "samples": "46025", "damp": "0.01000", "time": "14.976", "fwd_time": "2.720", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 48, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0004033617", "samples": "46025", "damp": "0.01000", "time": "3.316", "fwd_time": "0.659", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 48, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0001925261", "samples": "46025", "damp": "0.01000", "time": "3.380", "fwd_time": "0.477", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 48, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000145968", "samples": "46025", "damp": "0.01000", "time": "3.812", "fwd_time": "1.178", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 48, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0003387871", "samples": "46025", "damp": "0.01000", "time": "9.574", "fwd_time": "1.456", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 48, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0003879222", "samples": "46025", "damp": "0.01000", "time": "9.721", "fwd_time": "1.456", "(v)ram": "cuda 11.01G" }, { "process": "gptq", "layer": 48, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000278246", "samples": "46025", "damp": "0.01000", "time": "15.582", "fwd_time": "2.927", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 49, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0004812785", "samples": "46025", "damp": "0.01000", "time": "3.350", "fwd_time": "0.662", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 49, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0002058398", "samples": "46025", "damp": "0.01000", "time": "3.462", "fwd_time": "0.493", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 49, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000175645", "samples": "46025", "damp": "0.01000", "time": "3.177", "fwd_time": "1.102", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 49, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0003905665", "samples": "46025", "damp": "0.01000", "time": "9.249", "fwd_time": "1.436", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 49, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0003355557", "samples": "46025", "damp": "0.01000", "time": "9.542", "fwd_time": "1.436", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 49, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000318029", "samples": "46025", "damp": "0.01000", "time": "15.413", "fwd_time": "2.971", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 50, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0004514592", "samples": "46025", "damp": "0.01000", "time": "3.615", "fwd_time": "0.557", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 50, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0002119731", "samples": "46025", "damp": "0.01000", "time": "3.635", "fwd_time": "0.517", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 50, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000332887", "samples": "46025", "damp": "0.01000", "time": "6.183", "fwd_time": "1.036", "(v)ram": "cuda 11.07G" }, { "process": "gptq", "layer": 50, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0003830322", "samples": "46025", "damp": "0.01000", "time": "10.588", "fwd_time": "1.809", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 50, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0004344515", "samples": "46025", "damp": "0.01000", "time": "10.622", "fwd_time": "1.809", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 50, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000564420", "samples": "46025", "damp": "0.01000", "time": "17.167", "fwd_time": "3.148", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 51, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000520321", "samples": "46025", "damp": "0.01000", "time": "19.234", "fwd_time": "0.957", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 51, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0001140666", "samples": "46025", "damp": "0.01000", "time": "19.304", "fwd_time": "0.957", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 51, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0005503701", "samples": "46025", "damp": "0.01000", "time": "19.888", "fwd_time": "0.957", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 51, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0001009204", "samples": "46025", "damp": "0.01000", "time": "5.516", "fwd_time": "0.968", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 51, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0005437832", "samples": "46025", "damp": "0.01000", "time": "10.948", "fwd_time": "1.261", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 51, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0004680760", "samples": "46025", "damp": "0.01000", "time": "11.126", "fwd_time": "1.261", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 51, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000667506", "samples": "46025", "damp": "0.01000", "time": "18.175", "fwd_time": "2.646", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 52, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0006957806", "samples": "46025", "damp": "0.01000", "time": "4.101", "fwd_time": "0.669", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 52, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0003321920", "samples": "46025", "damp": "0.01000", "time": "4.831", "fwd_time": "0.493", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 52, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000290145", "samples": "46025", "damp": "0.01000", "time": "6.064", "fwd_time": "1.246", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 52, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0005294821", "samples": "46025", "damp": "0.01000", "time": "10.222", "fwd_time": "1.796", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 52, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0005367121", "samples": "46025", "damp": "0.01000", "time": "10.410", "fwd_time": "1.796", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 52, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000520043", "samples": "46025", "damp": "0.01000", "time": "20.238", "fwd_time": "2.913", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 53, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0006854756", "samples": "46025", "damp": "0.01000", "time": "4.503", "fwd_time": "0.624", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 53, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0003401085", "samples": "46025", "damp": "0.01000", "time": "5.666", "fwd_time": "0.589", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 53, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000246150", "samples": "46025", "damp": "0.01000", "time": "5.722", "fwd_time": "1.224", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 53, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0005381240", "samples": "46025", "damp": "0.01000", "time": "9.750", "fwd_time": "1.630", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 53, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0005620657", "samples": "46025", "damp": "0.01000", "time": "9.929", "fwd_time": "1.630", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 53, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000553583", "samples": "46025", "damp": "0.01000", "time": "26.157", "fwd_time": "3.074", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 54, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0006444547", "samples": "46025", "damp": "0.01000", "time": "7.576", "fwd_time": "0.643", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 54, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0003556014", "samples": "46025", "damp": "0.01000", "time": "5.881", "fwd_time": "0.647", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 54, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000318387", "samples": "46025", "damp": "0.01000", "time": "5.654", "fwd_time": "1.336", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 54, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0006923073", "samples": "46025", "damp": "0.01000", "time": "18.957", "fwd_time": "1.749", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 54, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0006238196", "samples": "46025", "damp": "0.01000", "time": "19.257", "fwd_time": "1.749", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 54, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000769702", "samples": "46025", "damp": "0.01000", "time": "25.285", "fwd_time": "2.906", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 55, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0001351974", "samples": "46025", "damp": "0.01000", "time": "23.919", "fwd_time": "1.051", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 55, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0006142842", "samples": "46025", "damp": "0.01000", "time": "27.177", "fwd_time": "1.051", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 55, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000571908", "samples": "46025", "damp": "0.01000", "time": "27.266", "fwd_time": "1.051", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 55, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000654712", "samples": "46025", "damp": "0.01000", "time": "6.597", "fwd_time": "1.001", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 55, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0006232404", "samples": "46025", "damp": "0.01000", "time": "14.274", "fwd_time": "1.704", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 55, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0006662700", "samples": "46025", "damp": "0.01000", "time": "14.517", "fwd_time": "1.704", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 55, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000553664", "samples": "46025", "damp": "0.01000", "time": "24.517", "fwd_time": "2.681", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 56, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0006978460", "samples": "46025", "damp": "0.01000", "time": "6.016", "fwd_time": "0.695", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 56, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0003547760", "samples": "46025", "damp": "0.01000", "time": "5.726", "fwd_time": "0.646", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 56, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000270589", "samples": "46025", "damp": "0.01000", "time": "5.025", "fwd_time": "1.312", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 56, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0006243683", "samples": "46025", "damp": "0.01000", "time": "11.159", "fwd_time": "1.558", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 56, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0006890403", "samples": "46025", "damp": "0.01000", "time": "11.266", "fwd_time": "1.558", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 56, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000400187", "samples": "46025", "damp": "0.01000", "time": "17.666", "fwd_time": "2.997", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 57, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0007181870", "samples": "46025", "damp": "0.01000", "time": "4.602", "fwd_time": "0.761", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 57, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0003695985", "samples": "46025", "damp": "0.01000", "time": "4.798", "fwd_time": "0.552", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 57, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000220725", "samples": "46025", "damp": "0.01000", "time": "4.982", "fwd_time": "1.165", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 57, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0007912585", "samples": "46025", "damp": "0.01000", "time": "10.215", "fwd_time": "1.649", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 57, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0007028669", "samples": "46025", "damp": "0.01000", "time": "10.432", "fwd_time": "1.649", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 57, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000420487", "samples": "46025", "damp": "0.01000", "time": "16.673", "fwd_time": "2.928", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 58, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0006966402", "samples": "46025", "damp": "0.01000", "time": "5.930", "fwd_time": "0.683", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 58, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0003850101", "samples": "46025", "damp": "0.01000", "time": "4.130", "fwd_time": "0.676", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 58, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000340235", "samples": "46025", "damp": "0.01000", "time": "5.907", "fwd_time": "1.196", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 58, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0008368557", "samples": "46025", "damp": "0.01000", "time": "12.306", "fwd_time": "1.612", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 58, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0009419893", "samples": "46025", "damp": "0.01000", "time": "12.361", "fwd_time": "1.612", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 58, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000704957", "samples": "46025", "damp": "0.01000", "time": "21.315", "fwd_time": "3.040", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 59, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000731669", "samples": "46025", "damp": "0.01000", "time": "20.285", "fwd_time": "0.912", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 59, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0002931732", "samples": "46025", "damp": "0.01000", "time": "20.341", "fwd_time": "0.912", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 59, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0007441575", "samples": "46025", "damp": "0.01000", "time": "20.695", "fwd_time": "0.912", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 59, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0001570079", "samples": "46025", "damp": "0.01000", "time": "4.908", "fwd_time": "0.860", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 59, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0009296697", "samples": "46025", "damp": "0.01000", "time": "12.360", "fwd_time": "1.450", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 59, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0008481844", "samples": "46025", "damp": "0.01000", "time": "12.453", "fwd_time": "1.450", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 59, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000995015", "samples": "46025", "damp": "0.01000", "time": "19.128", "fwd_time": "2.773", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 60, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0007892028", "samples": "46025", "damp": "0.01000", "time": "3.504", "fwd_time": "0.631", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 60, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0003849550", "samples": "46025", "damp": "0.01000", "time": "3.629", "fwd_time": "0.471", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 60, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000789163", "samples": "46025", "damp": "0.01000", "time": "4.657", "fwd_time": "1.147", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 60, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0008919395", "samples": "46025", "damp": "0.01000", "time": "11.949", "fwd_time": "1.628", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 60, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0008286461", "samples": "46025", "damp": "0.01000", "time": "12.032", "fwd_time": "1.628", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 60, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0001093913", "samples": "46025", "damp": "0.01000", "time": "17.382", "fwd_time": "2.872", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 61, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0006214758", "samples": "46025", "damp": "0.01000", "time": "3.093", "fwd_time": "0.572", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 61, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0003575272", "samples": "46025", "damp": "0.01000", "time": "3.513", "fwd_time": "0.470", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 61, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0001182015", "samples": "46025", "damp": "0.01000", "time": "4.195", "fwd_time": "1.099", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 61, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0008947431", "samples": "46025", "damp": "0.01000", "time": "11.354", "fwd_time": "1.580", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 61, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0009572177", "samples": "46025", "damp": "0.01000", "time": "11.722", "fwd_time": "1.580", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 61, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0001610767", "samples": "46025", "damp": "0.01000", "time": "16.241", "fwd_time": "2.968", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 62, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0006647699", "samples": "46025", "damp": "0.01000", "time": "3.924", "fwd_time": "0.568", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 62, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0003505199", "samples": "46025", "damp": "0.01000", "time": "3.596", "fwd_time": "0.500", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 62, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0003304372", "samples": "46025", "damp": "0.01000", "time": "4.076", "fwd_time": "1.137", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 62, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0008986733", "samples": "46025", "damp": "0.01000", "time": "11.234", "fwd_time": "1.443", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 62, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0008425839", "samples": "46025", "damp": "0.01000", "time": "11.425", "fwd_time": "1.443", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 62, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0002776341", "samples": "46025", "damp": "0.01000", "time": "14.449", "fwd_time": "2.875", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 63, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0002410854", "samples": "46025", "damp": "0.01000", "time": "20.372", "fwd_time": "0.966", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 63, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000650917", "samples": "46025", "damp": "0.01000", "time": "20.387", "fwd_time": "0.966", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 63, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0007221117", "samples": "46025", "damp": "0.01000", "time": "20.664", "fwd_time": "0.966", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 63, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0003938363", "samples": "46025", "damp": "0.01000", "time": "5.406", "fwd_time": "0.999", "(v)ram": "cuda 11.G" }, { "process": "gptq", "layer": 63, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0007631009", "samples": "46025", "damp": "0.01000", "time": "12.590", "fwd_time": "1.615", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 63, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0008420311", "samples": "46025", "damp": "0.01000", "time": "12.773", "fwd_time": "1.615", "(v)ram": "cuda 10.99G" }, { "process": "gptq", "layer": 63, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0006467385", "samples": "46025", "damp": "0.01000", "time": "25.297", "fwd_time": "2.746", "(v)ram": "cuda 11.G" } ] }