{ "gptq": [ { "process": "gptq", "layer": 0, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000002198", "samples": "46025", "damp": "0.01000", "time": "5.377", "fwd_time": "1.318", "(v)ram": "cuda 3.09G" }, { "process": "gptq", "layer": 0, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000001460", "samples": "46025", "damp": "0.01000", "time": "4.695", "fwd_time": "0.692", "(v)ram": "cuda 3.23G" }, { "process": "gptq", "layer": 0, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000003", "samples": "46025", "damp": "0.01000", "time": "6.195", "fwd_time": "1.351", "(v)ram": "cuda 3.4G" }, { "process": "gptq", "layer": 0, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000015", "samples": "46025", "damp": "0.01000", "time": "13.501", "fwd_time": "1.798", "(v)ram": "cuda 5.74G" }, { "process": "gptq", "layer": 0, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000014", "samples": "46025", "damp": "0.01000", "time": "13.704", "fwd_time": "1.798", "(v)ram": "cuda 5.74G" }, { "process": "gptq", "layer": 0, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000000", "samples": "46025", "damp": "0.01000", "time": "20.239", "fwd_time": "2.964", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 1, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000000232", "samples": "46025", "damp": "0.01000", "time": "5.174", "fwd_time": "0.670", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 1, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000142", "samples": "46025", "damp": "0.01000", "time": "4.748", "fwd_time": "0.561", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 1, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000000", "samples": "46025", "damp": "0.01000", "time": "6.685", "fwd_time": "1.216", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 1, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000031", "samples": "46025", "damp": "0.01000", "time": "12.231", "fwd_time": "1.642", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 1, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000029", "samples": "46025", "damp": "0.01000", "time": "12.318", "fwd_time": "1.642", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 1, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000000", "samples": "46025", "damp": "0.01000", "time": "19.494", "fwd_time": "3.073", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 2, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000000312", "samples": "46025", "damp": "0.01000", "time": "4.753", "fwd_time": "0.563", "(v)ram": "cuda 10.28G" }, { "process": "gptq", "layer": 2, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000189", "samples": "46025", "damp": "0.01000", "time": "5.763", "fwd_time": "0.604", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 2, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000001", "samples": "46025", "damp": "0.01000", "time": "5.519", "fwd_time": "1.313", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 2, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000055", "samples": "46025", "damp": "0.01000", "time": "12.124", "fwd_time": "1.575", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 2, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000058", "samples": "46025", "damp": "0.01000", "time": "12.154", "fwd_time": "1.575", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 2, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000000", "samples": "46025", "damp": "0.01000", "time": "18.463", "fwd_time": "3.031", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 3, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000000350", "samples": "46025", "damp": "0.01000", "time": "20.701", "fwd_time": "0.975", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 3, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000000356", "samples": "46025", "damp": "0.01000", "time": "21.157", "fwd_time": "0.975", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 3, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0000004884", "samples": "46025", "damp": "0.01000", "time": "21.183", "fwd_time": "0.975", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 3, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000001", "samples": "46025", "damp": "0.01000", "time": "4.814", "fwd_time": "1.009", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 3, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000065", "samples": "46025", "damp": "0.01000", "time": "12.560", "fwd_time": "1.296", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 3, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000067", "samples": "46025", "damp": "0.01000", "time": "12.638", "fwd_time": "1.296", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 3, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000000", "samples": "46025", "damp": "0.01000", "time": "16.877", "fwd_time": "2.651", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 4, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000000325", "samples": "46025", "damp": "0.01000", "time": "4.546", "fwd_time": "0.770", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 4, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000207", "samples": "46025", "damp": "0.01000", "time": "4.360", "fwd_time": "0.580", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 4, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000001", "samples": "46025", "damp": "0.01000", "time": "5.021", "fwd_time": "1.152", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 4, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000088", "samples": "46025", "damp": "0.01000", "time": "13.282", "fwd_time": "1.678", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 4, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000089", "samples": "46025", "damp": "0.01000", "time": "13.353", "fwd_time": "1.678", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 4, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000000", "samples": "46025", "damp": "0.01000", "time": "19.291", "fwd_time": "2.967", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 5, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000000460", "samples": "46025", "damp": "0.01000", "time": "4.507", "fwd_time": "0.505", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 5, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000299", "samples": "46025", "damp": "0.01000", "time": "4.332", "fwd_time": "0.614", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 5, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000001", "samples": "46025", "damp": "0.01000", "time": "6.202", "fwd_time": "1.194", "(v)ram": "cuda 10.27G" }, { "process": "gptq", "layer": 5, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000109", "samples": "46025", "damp": "0.01000", "time": "12.295", "fwd_time": "1.617", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 5, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000114", "samples": "46025", "damp": "0.01000", "time": "12.389", "fwd_time": "1.617", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 5, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000000", "samples": "46025", "damp": "0.01000", "time": "19.171", "fwd_time": "2.903", "(v)ram": "cuda 10.27G" }, { "process": "gptq", "layer": 6, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000000826", "samples": "46025", "damp": "0.01000", "time": "4.993", "fwd_time": "0.593", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 6, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000492", "samples": "46025", "damp": "0.01000", "time": "4.992", "fwd_time": "0.527", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 6, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000002", "samples": "46025", "damp": "0.01000", "time": "5.549", "fwd_time": "1.199", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 6, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000170", "samples": "46025", "damp": "0.01000", "time": "12.020", "fwd_time": "1.621", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 6, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000161", "samples": "46025", "damp": "0.01000", "time": "12.145", "fwd_time": "1.621", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 6, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000001", "samples": "46025", "damp": "0.01000", "time": "18.939", "fwd_time": "3.010", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 7, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000000287", "samples": "46025", "damp": "0.01000", "time": "21.889", "fwd_time": "1.073", "(v)ram": "cuda 10.27G" }, { "process": "gptq", "layer": 7, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000000298", "samples": "46025", "damp": "0.01000", "time": "21.983", "fwd_time": "1.073", "(v)ram": "cuda 10.27G" }, { "process": "gptq", "layer": 7, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0000003752", "samples": "46025", "damp": "0.01000", "time": "22.061", "fwd_time": "1.073", "(v)ram": "cuda 10.27G" }, { "process": "gptq", "layer": 7, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000002", "samples": "46025", "damp": "0.01000", "time": "5.742", "fwd_time": "0.728", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 7, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000179", "samples": "46025", "damp": "0.01000", "time": "13.788", "fwd_time": "1.371", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 7, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000189", "samples": "46025", "damp": "0.01000", "time": "13.980", "fwd_time": "1.371", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 7, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000001", "samples": "46025", "damp": "0.01000", "time": "18.749", "fwd_time": "2.717", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 8, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000000667", "samples": "46025", "damp": "0.01000", "time": "4.487", "fwd_time": "0.675", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 8, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000421", "samples": "46025", "damp": "0.01000", "time": "4.807", "fwd_time": "0.496", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 8, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000002", "samples": "46025", "damp": "0.01000", "time": "5.669", "fwd_time": "1.328", "(v)ram": "cuda 10.27G" }, { "process": "gptq", "layer": 8, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000208", "samples": "46025", "damp": "0.01000", "time": "13.354", "fwd_time": "1.679", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 8, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000198", "samples": "46025", "damp": "0.01000", "time": "13.527", "fwd_time": "1.679", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 8, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000001", "samples": "46025", "damp": "0.01000", "time": "20.125", "fwd_time": "2.845", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 9, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000000698", "samples": "46025", "damp": "0.01000", "time": "4.372", "fwd_time": "0.664", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 9, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000427", "samples": "46025", "damp": "0.01000", "time": "3.813", "fwd_time": "0.527", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 9, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000002", "samples": "46025", "damp": "0.01000", "time": "5.781", "fwd_time": "1.264", "(v)ram": "cuda 10.27G" }, { "process": "gptq", "layer": 9, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000216", "samples": "46025", "damp": "0.01000", "time": "12.341", "fwd_time": "1.702", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 9, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000207", "samples": "46025", "damp": "0.01000", "time": "12.426", "fwd_time": "1.702", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 9, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000001", "samples": "46025", "damp": "0.01000", "time": "18.252", "fwd_time": "2.982", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 10, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000000708", "samples": "46025", "damp": "0.01000", "time": "4.599", "fwd_time": "0.552", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 10, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000423", "samples": "46025", "damp": "0.01000", "time": "5.139", "fwd_time": "0.532", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 10, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000003", "samples": "46025", "damp": "0.01000", "time": "5.506", "fwd_time": "1.196", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 10, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000218", "samples": "46025", "damp": "0.01000", "time": "12.262", "fwd_time": "1.721", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 10, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000211", "samples": "46025", "damp": "0.01000", "time": "12.275", "fwd_time": "1.721", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 10, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000001", "samples": "46025", "damp": "0.01000", "time": "18.578", "fwd_time": "3.169", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 11, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000000245", "samples": "46025", "damp": "0.01000", "time": "21.183", "fwd_time": "1.018", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 11, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000000234", "samples": "46025", "damp": "0.01000", "time": "21.300", "fwd_time": "1.018", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 11, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0000003018", "samples": "46025", "damp": "0.01000", "time": "21.405", "fwd_time": "1.018", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 11, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000005", "samples": "46025", "damp": "0.01000", "time": "5.190", "fwd_time": "0.863", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 11, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000254", "samples": "46025", "damp": "0.01000", "time": "12.231", "fwd_time": "1.333", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 11, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000246", "samples": "46025", "damp": "0.01000", "time": "12.484", "fwd_time": "1.333", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 11, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000002", "samples": "46025", "damp": "0.01000", "time": "16.845", "fwd_time": "2.973", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 12, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000000921", "samples": "46025", "damp": "0.01000", "time": "4.463", "fwd_time": "0.620", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 12, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000547", "samples": "46025", "damp": "0.01000", "time": "4.422", "fwd_time": "0.537", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 12, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000004", "samples": "46025", "damp": "0.01000", "time": "5.017", "fwd_time": "1.156", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 12, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000266", "samples": "46025", "damp": "0.01000", "time": "11.961", "fwd_time": "1.669", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 12, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000257", "samples": "46025", "damp": "0.01000", "time": "12.063", "fwd_time": "1.669", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 12, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000002", "samples": "46025", "damp": "0.01000", "time": "16.732", "fwd_time": "2.987", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 13, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000000977", "samples": "46025", "damp": "0.01000", "time": "4.298", "fwd_time": "0.668", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 13, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000551", "samples": "46025", "damp": "0.01000", "time": "4.244", "fwd_time": "0.527", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 13, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000005", "samples": "46025", "damp": "0.01000", "time": "5.257", "fwd_time": "1.148", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 13, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000294", "samples": "46025", "damp": "0.01000", "time": "12.034", "fwd_time": "1.652", "(v)ram": "cuda 10.27G" }, { "process": "gptq", "layer": 13, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000283", "samples": "46025", "damp": "0.01000", "time": "12.162", "fwd_time": "1.652", "(v)ram": "cuda 10.27G" }, { "process": "gptq", "layer": 13, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000002", "samples": "46025", "damp": "0.01000", "time": "16.875", "fwd_time": "3.033", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 14, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000001089", "samples": "46025", "damp": "0.01000", "time": "4.656", "fwd_time": "0.608", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 14, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000639", "samples": "46025", "damp": "0.01000", "time": "4.231", "fwd_time": "0.492", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 14, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000009", "samples": "46025", "damp": "0.01000", "time": "5.907", "fwd_time": "1.224", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 14, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000310", "samples": "46025", "damp": "0.01000", "time": "12.939", "fwd_time": "1.758", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 14, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000321", "samples": "46025", "damp": "0.01000", "time": "12.969", "fwd_time": "1.758", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 14, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000003", "samples": "46025", "damp": "0.01000", "time": "18.205", "fwd_time": "2.985", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 15, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000000206", "samples": "46025", "damp": "0.01000", "time": "21.730", "fwd_time": "1.050", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 15, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000000205", "samples": "46025", "damp": "0.01000", "time": "21.978", "fwd_time": "1.050", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 15, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0000002372", "samples": "46025", "damp": "0.01000", "time": "22.037", "fwd_time": "1.050", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 15, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000007", "samples": "46025", "damp": "0.01000", "time": "5.302", "fwd_time": "0.932", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 15, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000378", "samples": "46025", "damp": "0.01000", "time": "12.070", "fwd_time": "1.327", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 15, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000387", "samples": "46025", "damp": "0.01000", "time": "12.119", "fwd_time": "1.327", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 15, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000004", "samples": "46025", "damp": "0.01000", "time": "17.182", "fwd_time": "2.726", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 16, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000001167", "samples": "46025", "damp": "0.01000", "time": "5.359", "fwd_time": "0.554", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 16, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000663", "samples": "46025", "damp": "0.01000", "time": "4.514", "fwd_time": "0.508", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 16, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000008", "samples": "46025", "damp": "0.01000", "time": "5.679", "fwd_time": "1.105", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 16, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000427", "samples": "46025", "damp": "0.01000", "time": "12.436", "fwd_time": "1.649", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 16, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000438", "samples": "46025", "damp": "0.01000", "time": "12.505", "fwd_time": "1.649", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 16, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000005", "samples": "46025", "damp": "0.01000", "time": "18.194", "fwd_time": "3.013", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 17, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000001348", "samples": "46025", "damp": "0.01000", "time": "4.817", "fwd_time": "0.659", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 17, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000691", "samples": "46025", "damp": "0.01000", "time": "4.486", "fwd_time": "0.564", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 17, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000011", "samples": "46025", "damp": "0.01000", "time": "5.290", "fwd_time": "1.147", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 17, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000520", "samples": "46025", "damp": "0.01000", "time": "12.154", "fwd_time": "1.638", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 17, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000507", "samples": "46025", "damp": "0.01000", "time": "12.179", "fwd_time": "1.638", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 17, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000006", "samples": "46025", "damp": "0.01000", "time": "16.961", "fwd_time": "2.976", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 18, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000001452", "samples": "46025", "damp": "0.01000", "time": "4.551", "fwd_time": "0.539", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 18, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000807", "samples": "46025", "damp": "0.01000", "time": "4.029", "fwd_time": "0.533", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 18, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000018", "samples": "46025", "damp": "0.01000", "time": "4.044", "fwd_time": "1.106", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 18, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000677", "samples": "46025", "damp": "0.01000", "time": "9.737", "fwd_time": "1.552", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 18, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000000657", "samples": "46025", "damp": "0.01000", "time": "9.758", "fwd_time": "1.552", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 18, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000013", "samples": "46025", "damp": "0.01000", "time": "12.532", "fwd_time": "2.982", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 19, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000000241", "samples": "46025", "damp": "0.01000", "time": "16.470", "fwd_time": "1.068", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 19, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0000002577", "samples": "46025", "damp": "0.01000", "time": "16.552", "fwd_time": "1.068", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 19, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000000219", "samples": "46025", "damp": "0.01000", "time": "16.614", "fwd_time": "1.068", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 19, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000043", "samples": "46025", "damp": "0.01000", "time": "3.907", "fwd_time": "0.878", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 19, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000001000", "samples": "46025", "damp": "0.01000", "time": "9.709", "fwd_time": "1.285", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 19, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000001033", "samples": "46025", "damp": "0.01000", "time": "9.791", "fwd_time": "1.285", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 19, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000021", "samples": "46025", "damp": "0.01000", "time": "12.486", "fwd_time": "2.558", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 20, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000002986", "samples": "46025", "damp": "0.01000", "time": "3.639", "fwd_time": "0.545", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 20, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000001705", "samples": "46025", "damp": "0.01000", "time": "3.215", "fwd_time": "0.462", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 20, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000039", "samples": "46025", "damp": "0.01000", "time": "4.100", "fwd_time": "1.129", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 20, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000001336", "samples": "46025", "damp": "0.01000", "time": "9.252", "fwd_time": "1.576", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 20, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000001270", "samples": "46025", "damp": "0.01000", "time": "9.406", "fwd_time": "1.576", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 20, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000030", "samples": "46025", "damp": "0.01000", "time": "13.758", "fwd_time": "2.898", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 21, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000003611", "samples": "46025", "damp": "0.01000", "time": "3.350", "fwd_time": "0.608", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 21, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000002334", "samples": "46025", "damp": "0.01000", "time": "3.245", "fwd_time": "0.451", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 21, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000043", "samples": "46025", "damp": "0.01000", "time": "4.149", "fwd_time": "1.126", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 21, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000001441", "samples": "46025", "damp": "0.01000", "time": "9.452", "fwd_time": "1.559", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 21, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000001514", "samples": "46025", "damp": "0.01000", "time": "9.466", "fwd_time": "1.559", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 21, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000027", "samples": "46025", "damp": "0.01000", "time": "12.782", "fwd_time": "2.956", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 22, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000003188", "samples": "46025", "damp": "0.01000", "time": "3.158", "fwd_time": "0.489", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 22, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000001974", "samples": "46025", "damp": "0.01000", "time": "3.508", "fwd_time": "0.430", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 22, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000033", "samples": "46025", "damp": "0.01000", "time": "3.232", "fwd_time": "1.098", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 22, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000001838", "samples": "46025", "damp": "0.01000", "time": "9.367", "fwd_time": "1.531", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 22, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000001667", "samples": "46025", "damp": "0.01000", "time": "9.559", "fwd_time": "1.531", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 22, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000032", "samples": "46025", "damp": "0.01000", "time": "12.387", "fwd_time": "2.887", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 23, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000000380", "samples": "46025", "damp": "0.01000", "time": "17.308", "fwd_time": "0.935", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 23, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000000337", "samples": "46025", "damp": "0.01000", "time": "17.398", "fwd_time": "0.935", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 23, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0000003773", "samples": "46025", "damp": "0.01000", "time": "17.452", "fwd_time": "0.935", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 23, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000057", "samples": "46025", "damp": "0.01000", "time": "3.803", "fwd_time": "0.759", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 23, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000002193", "samples": "46025", "damp": "0.01000", "time": "9.443", "fwd_time": "1.243", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 23, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000002067", "samples": "46025", "damp": "0.01000", "time": "9.546", "fwd_time": "1.243", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 23, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000040", "samples": "46025", "damp": "0.01000", "time": "12.933", "fwd_time": "2.629", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 24, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000003341", "samples": "46025", "damp": "0.01000", "time": "3.280", "fwd_time": "0.583", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 24, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000001923", "samples": "46025", "damp": "0.01000", "time": "3.503", "fwd_time": "0.498", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 24, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000061", "samples": "46025", "damp": "0.01000", "time": "3.887", "fwd_time": "1.150", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 24, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000002399", "samples": "46025", "damp": "0.01000", "time": "9.274", "fwd_time": "1.538", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 24, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000002550", "samples": "46025", "damp": "0.01000", "time": "9.325", "fwd_time": "1.538", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 24, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000052", "samples": "46025", "damp": "0.01000", "time": "12.417", "fwd_time": "2.879", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 25, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000003624", "samples": "46025", "damp": "0.01000", "time": "3.758", "fwd_time": "0.527", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 25, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000001963", "samples": "46025", "damp": "0.01000", "time": "3.527", "fwd_time": "0.458", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 25, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000066", "samples": "46025", "damp": "0.01000", "time": "3.732", "fwd_time": "1.167", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 25, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000002479", "samples": "46025", "damp": "0.01000", "time": "9.931", "fwd_time": "1.587", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 25, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000002444", "samples": "46025", "damp": "0.01000", "time": "9.970", "fwd_time": "1.587", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 25, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000058", "samples": "46025", "damp": "0.01000", "time": "13.005", "fwd_time": "2.865", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 26, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000003353", "samples": "46025", "damp": "0.01000", "time": "3.576", "fwd_time": "0.458", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 26, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000001753", "samples": "46025", "damp": "0.01000", "time": "3.549", "fwd_time": "0.448", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 26, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000108", "samples": "46025", "damp": "0.01000", "time": "3.651", "fwd_time": "1.059", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 26, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000002477", "samples": "46025", "damp": "0.01000", "time": "9.526", "fwd_time": "1.551", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 26, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000002560", "samples": "46025", "damp": "0.01000", "time": "9.548", "fwd_time": "1.551", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 26, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000076", "samples": "46025", "damp": "0.01000", "time": "12.613", "fwd_time": "2.968", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 27, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000000394", "samples": "46025", "damp": "0.01000", "time": "16.255", "fwd_time": "0.861", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 27, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0000004256", "samples": "46025", "damp": "0.01000", "time": "16.339", "fwd_time": "0.861", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 27, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000000381", "samples": "46025", "damp": "0.01000", "time": "16.411", "fwd_time": "0.861", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 27, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000194", "samples": "46025", "damp": "0.01000", "time": "3.947", "fwd_time": "0.949", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 27, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000003018", "samples": "46025", "damp": "0.01000", "time": "9.407", "fwd_time": "1.232", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 27, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000003227", "samples": "46025", "damp": "0.01000", "time": "9.585", "fwd_time": "1.232", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 27, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000110", "samples": "46025", "damp": "0.01000", "time": "12.829", "fwd_time": "2.587", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 28, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000004239", "samples": "46025", "damp": "0.01000", "time": "3.392", "fwd_time": "0.458", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 28, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000002035", "samples": "46025", "damp": "0.01000", "time": "3.399", "fwd_time": "0.451", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 28, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000138", "samples": "46025", "damp": "0.01000", "time": "4.434", "fwd_time": "1.119", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 28, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000003153", "samples": "46025", "damp": "0.01000", "time": "9.464", "fwd_time": "1.508", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 28, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000003430", "samples": "46025", "damp": "0.01000", "time": "9.560", "fwd_time": "1.508", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 28, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000119", "samples": "46025", "damp": "0.01000", "time": "12.849", "fwd_time": "2.913", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 29, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000004943", "samples": "46025", "damp": "0.01000", "time": "3.364", "fwd_time": "0.534", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 29, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000002298", "samples": "46025", "damp": "0.01000", "time": "3.676", "fwd_time": "0.440", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 29, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000148", "samples": "46025", "damp": "0.01000", "time": "3.977", "fwd_time": "1.104", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 29, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000003388", "samples": "46025", "damp": "0.01000", "time": "10.056", "fwd_time": "1.574", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 29, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000003703", "samples": "46025", "damp": "0.01000", "time": "10.135", "fwd_time": "1.574", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 29, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000144", "samples": "46025", "damp": "0.01000", "time": "12.140", "fwd_time": "2.984", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 30, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000005305", "samples": "46025", "damp": "0.01000", "time": "3.476", "fwd_time": "0.469", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 30, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000002529", "samples": "46025", "damp": "0.01000", "time": "3.431", "fwd_time": "0.465", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 30, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000236", "samples": "46025", "damp": "0.01000", "time": "3.284", "fwd_time": "1.131", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 30, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000004074", "samples": "46025", "damp": "0.01000", "time": "9.329", "fwd_time": "1.619", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 30, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000003650", "samples": "46025", "damp": "0.01000", "time": "9.464", "fwd_time": "1.619", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 30, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000196", "samples": "46025", "damp": "0.01000", "time": "13.232", "fwd_time": "2.920", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 31, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000000715", "samples": "46025", "damp": "0.01000", "time": "17.616", "fwd_time": "0.925", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 31, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000000576", "samples": "46025", "damp": "0.01000", "time": "17.727", "fwd_time": "0.925", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 31, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0000005921", "samples": "46025", "damp": "0.01000", "time": "17.797", "fwd_time": "0.925", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 31, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000389", "samples": "46025", "damp": "0.01000", "time": "3.728", "fwd_time": "0.790", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 31, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000004898", "samples": "46025", "damp": "0.01000", "time": "9.668", "fwd_time": "1.222", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 31, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000004302", "samples": "46025", "damp": "0.01000", "time": "9.683", "fwd_time": "1.222", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 31, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000262", "samples": "46025", "damp": "0.01000", "time": "11.531", "fwd_time": "2.614", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 32, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000006827", "samples": "46025", "damp": "0.01000", "time": "3.238", "fwd_time": "0.469", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 32, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000003029", "samples": "46025", "damp": "0.01000", "time": "3.109", "fwd_time": "0.437", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 32, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000296", "samples": "46025", "damp": "0.01000", "time": "3.967", "fwd_time": "1.119", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 32, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000005191", "samples": "46025", "damp": "0.01000", "time": "9.687", "fwd_time": "1.501", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 32, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000005941", "samples": "46025", "damp": "0.01000", "time": "9.793", "fwd_time": "1.501", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 32, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000317", "samples": "46025", "damp": "0.01000", "time": "12.851", "fwd_time": "2.953", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 33, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000008830", "samples": "46025", "damp": "0.01000", "time": "3.072", "fwd_time": "0.581", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 33, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000003672", "samples": "46025", "damp": "0.01000", "time": "3.123", "fwd_time": "0.451", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 33, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000394", "samples": "46025", "damp": "0.01000", "time": "4.007", "fwd_time": "1.152", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 33, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000006985", "samples": "46025", "damp": "0.01000", "time": "9.316", "fwd_time": "1.523", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 33, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000006026", "samples": "46025", "damp": "0.01000", "time": "9.476", "fwd_time": "1.523", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 33, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000347", "samples": "46025", "damp": "0.01000", "time": "12.214", "fwd_time": "2.849", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 34, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000009351", "samples": "46025", "damp": "0.01000", "time": "3.536", "fwd_time": "0.502", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 34, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000004260", "samples": "46025", "damp": "0.01000", "time": "3.552", "fwd_time": "0.471", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 34, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000392", "samples": "46025", "damp": "0.01000", "time": "3.119", "fwd_time": "1.106", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 34, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000006519", "samples": "46025", "damp": "0.01000", "time": "9.583", "fwd_time": "1.600", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 34, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000007755", "samples": "46025", "damp": "0.01000", "time": "9.700", "fwd_time": "1.600", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 34, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000468", "samples": "46025", "damp": "0.01000", "time": "13.807", "fwd_time": "2.936", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 35, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0000010194", "samples": "46025", "damp": "0.01000", "time": "16.606", "fwd_time": "0.836", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 35, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000001506", "samples": "46025", "damp": "0.01000", "time": "16.782", "fwd_time": "0.836", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 35, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000000975", "samples": "46025", "damp": "0.01000", "time": "16.796", "fwd_time": "0.836", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 35, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000835", "samples": "46025", "damp": "0.01000", "time": "4.145", "fwd_time": "0.984", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 35, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000007342", "samples": "46025", "damp": "0.01000", "time": "8.782", "fwd_time": "1.239", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 35, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000008365", "samples": "46025", "damp": "0.01000", "time": "8.850", "fwd_time": "1.239", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 35, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000430", "samples": "46025", "damp": "0.01000", "time": "13.331", "fwd_time": "2.647", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 36, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000012972", "samples": "46025", "damp": "0.01000", "time": "3.646", "fwd_time": "0.613", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 36, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000006611", "samples": "46025", "damp": "0.01000", "time": "3.344", "fwd_time": "0.482", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 36, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000215", "samples": "46025", "damp": "0.01000", "time": "4.105", "fwd_time": "1.068", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 36, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000008026", "samples": "46025", "damp": "0.01000", "time": "9.526", "fwd_time": "1.582", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 36, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000008119", "samples": "46025", "damp": "0.01000", "time": "9.673", "fwd_time": "1.582", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 36, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000333", "samples": "46025", "damp": "0.01000", "time": "12.206", "fwd_time": "2.986", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 37, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000013463", "samples": "46025", "damp": "0.01000", "time": "3.929", "fwd_time": "0.549", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 37, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000007766", "samples": "46025", "damp": "0.01000", "time": "3.282", "fwd_time": "0.479", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 37, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000233", "samples": "46025", "damp": "0.01000", "time": "3.930", "fwd_time": "1.099", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 37, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000007582", "samples": "46025", "damp": "0.01000", "time": "9.716", "fwd_time": "1.562", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 37, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000007686", "samples": "46025", "damp": "0.01000", "time": "9.840", "fwd_time": "1.562", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 37, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000294", "samples": "46025", "damp": "0.01000", "time": "13.244", "fwd_time": "2.986", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 38, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000012186", "samples": "46025", "damp": "0.01000", "time": "3.387", "fwd_time": "0.520", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 38, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000007188", "samples": "46025", "damp": "0.01000", "time": "3.432", "fwd_time": "0.439", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 38, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000170", "samples": "46025", "damp": "0.01000", "time": "3.992", "fwd_time": "1.152", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 38, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000010018", "samples": "46025", "damp": "0.01000", "time": "9.501", "fwd_time": "1.637", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 38, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000008628", "samples": "46025", "damp": "0.01000", "time": "9.615", "fwd_time": "1.637", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 38, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000310", "samples": "46025", "damp": "0.01000", "time": "12.246", "fwd_time": "3.062", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 39, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0000010371", "samples": "46025", "damp": "0.01000", "time": "16.393", "fwd_time": "0.923", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 39, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000001267", "samples": "46025", "damp": "0.01000", "time": "16.579", "fwd_time": "0.923", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 39, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000000942", "samples": "46025", "damp": "0.01000", "time": "16.588", "fwd_time": "0.923", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 39, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000506", "samples": "46025", "damp": "0.01000", "time": "3.891", "fwd_time": "0.937", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 39, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000008802", "samples": "46025", "damp": "0.01000", "time": "9.620", "fwd_time": "1.247", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 39, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000008064", "samples": "46025", "damp": "0.01000", "time": "9.729", "fwd_time": "1.247", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 39, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000256", "samples": "46025", "damp": "0.01000", "time": "11.908", "fwd_time": "2.538", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 40, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000010391", "samples": "46025", "damp": "0.01000", "time": "3.460", "fwd_time": "0.573", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 40, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000005809", "samples": "46025", "damp": "0.01000", "time": "3.411", "fwd_time": "0.473", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 40, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000222", "samples": "46025", "damp": "0.01000", "time": "3.900", "fwd_time": "1.155", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 40, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000008590", "samples": "46025", "damp": "0.01000", "time": "8.699", "fwd_time": "1.487", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 40, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000009547", "samples": "46025", "damp": "0.01000", "time": "8.758", "fwd_time": "1.487", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 40, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000294", "samples": "46025", "damp": "0.01000", "time": "13.123", "fwd_time": "2.916", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 41, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000009743", "samples": "46025", "damp": "0.01000", "time": "3.609", "fwd_time": "0.590", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 41, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000005433", "samples": "46025", "damp": "0.01000", "time": "3.837", "fwd_time": "0.477", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 41, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000172", "samples": "46025", "damp": "0.01000", "time": "4.027", "fwd_time": "1.164", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 41, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000007994", "samples": "46025", "damp": "0.01000", "time": "9.512", "fwd_time": "1.627", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 41, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000007938", "samples": "46025", "damp": "0.01000", "time": "9.554", "fwd_time": "1.627", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 41, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000317", "samples": "46025", "damp": "0.01000", "time": "12.699", "fwd_time": "2.803", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 42, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000008953", "samples": "46025", "damp": "0.01000", "time": "3.316", "fwd_time": "0.643", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 42, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000004829", "samples": "46025", "damp": "0.01000", "time": "3.443", "fwd_time": "0.466", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 42, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000321", "samples": "46025", "damp": "0.01000", "time": "3.088", "fwd_time": "1.124", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 42, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000007776", "samples": "46025", "damp": "0.01000", "time": "9.782", "fwd_time": "1.475", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 42, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000007318", "samples": "46025", "damp": "0.01000", "time": "9.886", "fwd_time": "1.475", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 42, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000401", "samples": "46025", "damp": "0.01000", "time": "13.176", "fwd_time": "2.912", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 43, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000000962", "samples": "46025", "damp": "0.01000", "time": "17.097", "fwd_time": "0.937", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 43, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000001331", "samples": "46025", "damp": "0.01000", "time": "17.219", "fwd_time": "0.937", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 43, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0000010174", "samples": "46025", "damp": "0.01000", "time": "17.256", "fwd_time": "0.937", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 43, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000971", "samples": "46025", "damp": "0.01000", "time": "4.054", "fwd_time": "0.901", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 43, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000009051", "samples": "46025", "damp": "0.01000", "time": "9.146", "fwd_time": "1.244", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 43, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000008180", "samples": "46025", "damp": "0.01000", "time": "9.161", "fwd_time": "1.244", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 43, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000488", "samples": "46025", "damp": "0.01000", "time": "13.202", "fwd_time": "2.664", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 44, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000010221", "samples": "46025", "damp": "0.01000", "time": "3.486", "fwd_time": "0.591", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 44, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000005039", "samples": "46025", "damp": "0.01000", "time": "3.247", "fwd_time": "0.476", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 44, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000372", "samples": "46025", "damp": "0.01000", "time": "3.824", "fwd_time": "1.150", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 44, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000008437", "samples": "46025", "damp": "0.01000", "time": "9.500", "fwd_time": "1.528", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 44, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000009420", "samples": "46025", "damp": "0.01000", "time": "9.579", "fwd_time": "1.528", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 44, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000517", "samples": "46025", "damp": "0.01000", "time": "12.511", "fwd_time": "2.926", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 45, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000011050", "samples": "46025", "damp": "0.01000", "time": "3.483", "fwd_time": "0.499", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 45, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000005391", "samples": "46025", "damp": "0.01000", "time": "3.602", "fwd_time": "0.479", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 45, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000339", "samples": "46025", "damp": "0.01000", "time": "3.482", "fwd_time": "1.169", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 45, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000009696", "samples": "46025", "damp": "0.01000", "time": "9.559", "fwd_time": "1.632", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 45, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000008648", "samples": "46025", "damp": "0.01000", "time": "9.718", "fwd_time": "1.632", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 45, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000567", "samples": "46025", "damp": "0.01000", "time": "12.976", "fwd_time": "2.963", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 46, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000011622", "samples": "46025", "damp": "0.01000", "time": "3.285", "fwd_time": "0.536", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 46, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000005731", "samples": "46025", "damp": "0.01000", "time": "3.175", "fwd_time": "0.460", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 46, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000547", "samples": "46025", "damp": "0.01000", "time": "3.914", "fwd_time": "1.149", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 46, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000010052", "samples": "46025", "damp": "0.01000", "time": "9.590", "fwd_time": "1.504", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 46, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000008772", "samples": "46025", "damp": "0.01000", "time": "9.626", "fwd_time": "1.504", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 46, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000656", "samples": "46025", "damp": "0.01000", "time": "12.072", "fwd_time": "2.943", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 47, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000001300", "samples": "46025", "damp": "0.01000", "time": "16.719", "fwd_time": "1.040", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 47, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0000013534", "samples": "46025", "damp": "0.01000", "time": "16.835", "fwd_time": "1.040", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 47, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000002258", "samples": "46025", "damp": "0.01000", "time": "16.928", "fwd_time": "1.040", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 47, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000001323", "samples": "46025", "damp": "0.01000", "time": "4.251", "fwd_time": "0.933", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 47, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000011988", "samples": "46025", "damp": "0.01000", "time": "9.384", "fwd_time": "1.225", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 47, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000010508", "samples": "46025", "damp": "0.01000", "time": "9.433", "fwd_time": "1.225", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 47, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000847", "samples": "46025", "damp": "0.01000", "time": "11.963", "fwd_time": "2.600", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 48, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000013732", "samples": "46025", "damp": "0.01000", "time": "3.573", "fwd_time": "0.504", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 48, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000006551", "samples": "46025", "damp": "0.01000", "time": "3.460", "fwd_time": "0.454", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 48, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000502", "samples": "46025", "damp": "0.01000", "time": "3.859", "fwd_time": "1.131", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 48, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000013195", "samples": "46025", "damp": "0.01000", "time": "8.508", "fwd_time": "1.514", "(v)ram": "cuda 10.27G" }, { "process": "gptq", "layer": 48, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000011529", "samples": "46025", "damp": "0.01000", "time": "8.657", "fwd_time": "1.514", "(v)ram": "cuda 10.27G" }, { "process": "gptq", "layer": 48, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000000941", "samples": "46025", "damp": "0.01000", "time": "12.999", "fwd_time": "2.934", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 49, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000016388", "samples": "46025", "damp": "0.01000", "time": "3.325", "fwd_time": "0.308", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 49, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000007001", "samples": "46025", "damp": "0.01000", "time": "3.345", "fwd_time": "0.449", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 49, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000598", "samples": "46025", "damp": "0.01000", "time": "3.981", "fwd_time": "1.114", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 49, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000013280", "samples": "46025", "damp": "0.01000", "time": "9.674", "fwd_time": "1.524", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 49, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000011409", "samples": "46025", "damp": "0.01000", "time": "9.758", "fwd_time": "1.524", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 49, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000001075", "samples": "46025", "damp": "0.01000", "time": "12.236", "fwd_time": "2.974", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 50, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000015370", "samples": "46025", "damp": "0.01000", "time": "3.898", "fwd_time": "0.607", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 50, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000007211", "samples": "46025", "damp": "0.01000", "time": "3.600", "fwd_time": "0.499", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 50, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000001151", "samples": "46025", "damp": "0.01000", "time": "3.682", "fwd_time": "1.131", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 50, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000014778", "samples": "46025", "damp": "0.01000", "time": "8.653", "fwd_time": "1.523", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 50, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000013012", "samples": "46025", "damp": "0.01000", "time": "8.754", "fwd_time": "1.523", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 50, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000001920", "samples": "46025", "damp": "0.01000", "time": "12.883", "fwd_time": "2.941", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 51, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000001766", "samples": "46025", "damp": "0.01000", "time": "17.233", "fwd_time": "0.818", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 51, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0000018684", "samples": "46025", "damp": "0.01000", "time": "17.380", "fwd_time": "0.818", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 51, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000003890", "samples": "46025", "damp": "0.01000", "time": "17.431", "fwd_time": "0.818", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 51, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000003470", "samples": "46025", "damp": "0.01000", "time": "4.070", "fwd_time": "0.779", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 51, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000018500", "samples": "46025", "damp": "0.01000", "time": "8.796", "fwd_time": "1.196", "(v)ram": "cuda 10.24G" }, { "process": "gptq", "layer": 51, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000015913", "samples": "46025", "damp": "0.01000", "time": "8.895", "fwd_time": "1.196", "(v)ram": "cuda 10.24G" }, { "process": "gptq", "layer": 51, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000002265", "samples": "46025", "damp": "0.01000", "time": "13.551", "fwd_time": "2.551", "(v)ram": "cuda 10.24G" }, { "process": "gptq", "layer": 52, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000023667", "samples": "46025", "damp": "0.01000", "time": "3.526", "fwd_time": "0.582", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 52, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000011304", "samples": "46025", "damp": "0.01000", "time": "2.789", "fwd_time": "0.500", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 52, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000001028", "samples": "46025", "damp": "0.01000", "time": "3.957", "fwd_time": "1.110", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 52, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000017980", "samples": "46025", "damp": "0.01000", "time": "9.819", "fwd_time": "1.633", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 52, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000018247", "samples": "46025", "damp": "0.01000", "time": "9.838", "fwd_time": "1.633", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 52, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000001765", "samples": "46025", "damp": "0.01000", "time": "12.406", "fwd_time": "2.989", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 53, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000023309", "samples": "46025", "damp": "0.01000", "time": "3.671", "fwd_time": "0.609", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 53, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000011562", "samples": "46025", "damp": "0.01000", "time": "3.739", "fwd_time": "0.474", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 53, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000849", "samples": "46025", "damp": "0.01000", "time": "4.236", "fwd_time": "1.108", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 53, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000018306", "samples": "46025", "damp": "0.01000", "time": "8.754", "fwd_time": "1.579", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 53, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000019120", "samples": "46025", "damp": "0.01000", "time": "8.797", "fwd_time": "1.579", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 53, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000001880", "samples": "46025", "damp": "0.01000", "time": "13.145", "fwd_time": "2.996", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 54, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000021940", "samples": "46025", "damp": "0.01000", "time": "3.558", "fwd_time": "0.517", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 54, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000012096", "samples": "46025", "damp": "0.01000", "time": "3.563", "fwd_time": "0.476", "(v)ram": "cuda 10.24G" }, { "process": "gptq", "layer": 54, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000001140", "samples": "46025", "damp": "0.01000", "time": "4.119", "fwd_time": "1.165", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 54, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000021231", "samples": "46025", "damp": "0.01000", "time": "9.270", "fwd_time": "1.565", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 54, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000023554", "samples": "46025", "damp": "0.01000", "time": "9.464", "fwd_time": "1.565", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 54, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000002620", "samples": "46025", "damp": "0.01000", "time": "12.893", "fwd_time": "2.957", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 55, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000004638", "samples": "46025", "damp": "0.01000", "time": "16.325", "fwd_time": "0.994", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 55, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000001931", "samples": "46025", "damp": "0.01000", "time": "16.390", "fwd_time": "0.994", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 55, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0000020884", "samples": "46025", "damp": "0.01000", "time": "16.392", "fwd_time": "0.994", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 55, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000002251", "samples": "46025", "damp": "0.01000", "time": "4.301", "fwd_time": "0.843", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 55, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000021248", "samples": "46025", "damp": "0.01000", "time": "9.540", "fwd_time": "1.259", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 55, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000022698", "samples": "46025", "damp": "0.01000", "time": "9.572", "fwd_time": "1.259", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 55, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000001896", "samples": "46025", "damp": "0.01000", "time": "12.508", "fwd_time": "2.556", "(v)ram": "cuda 10.24G" }, { "process": "gptq", "layer": 56, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000023759", "samples": "46025", "damp": "0.01000", "time": "4.060", "fwd_time": "0.549", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 56, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000012079", "samples": "46025", "damp": "0.01000", "time": "3.101", "fwd_time": "0.502", "(v)ram": "cuda 10.24G" }, { "process": "gptq", "layer": 56, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000947", "samples": "46025", "damp": "0.01000", "time": "4.286", "fwd_time": "1.096", "(v)ram": "cuda 10.24G" }, { "process": "gptq", "layer": 56, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000021268", "samples": "46025", "damp": "0.01000", "time": "8.696", "fwd_time": "1.599", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 56, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000023475", "samples": "46025", "damp": "0.01000", "time": "8.856", "fwd_time": "1.599", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 56, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000001357", "samples": "46025", "damp": "0.01000", "time": "12.742", "fwd_time": "2.943", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 57, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000024475", "samples": "46025", "damp": "0.01000", "time": "3.477", "fwd_time": "0.538", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 57, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000012589", "samples": "46025", "damp": "0.01000", "time": "3.467", "fwd_time": "0.479", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 57, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000000759", "samples": "46025", "damp": "0.01000", "time": "3.935", "fwd_time": "1.154", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 57, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000026967", "samples": "46025", "damp": "0.01000", "time": "9.479", "fwd_time": "1.599", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 57, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000023949", "samples": "46025", "damp": "0.01000", "time": "9.498", "fwd_time": "1.599", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 57, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000001424", "samples": "46025", "damp": "0.01000", "time": "12.715", "fwd_time": "2.997", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 58, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000023730", "samples": "46025", "damp": "0.01000", "time": "3.283", "fwd_time": "0.509", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 58, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000013110", "samples": "46025", "damp": "0.01000", "time": "3.432", "fwd_time": "0.458", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 58, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000001203", "samples": "46025", "damp": "0.01000", "time": "4.052", "fwd_time": "1.134", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 58, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000028532", "samples": "46025", "damp": "0.01000", "time": "8.292", "fwd_time": "1.615", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 58, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000032105", "samples": "46025", "damp": "0.01000", "time": "8.418", "fwd_time": "1.615", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 58, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000002417", "samples": "46025", "damp": "0.01000", "time": "11.713", "fwd_time": "2.948", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 59, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0000025304", "samples": "46025", "damp": "0.01000", "time": "15.057", "fwd_time": "0.889", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 59, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000002481", "samples": "46025", "damp": "0.01000", "time": "15.179", "fwd_time": "0.889", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 59, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000010416", "samples": "46025", "damp": "0.01000", "time": "15.188", "fwd_time": "0.889", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 59, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000005371", "samples": "46025", "damp": "0.01000", "time": "3.875", "fwd_time": "0.861", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 59, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000031702", "samples": "46025", "damp": "0.01000", "time": "9.007", "fwd_time": "1.254", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 59, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000028934", "samples": "46025", "damp": "0.01000", "time": "9.103", "fwd_time": "1.254", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 59, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000003399", "samples": "46025", "damp": "0.01000", "time": "11.054", "fwd_time": "2.581", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 60, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000026900", "samples": "46025", "damp": "0.01000", "time": "3.293", "fwd_time": "0.368", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 60, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000013115", "samples": "46025", "damp": "0.01000", "time": "3.036", "fwd_time": "0.406", "(v)ram": "cuda 10.26G" }, { "process": "gptq", "layer": 60, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000002739", "samples": "46025", "damp": "0.01000", "time": "3.622", "fwd_time": "1.101", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 60, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000028263", "samples": "46025", "damp": "0.01000", "time": "8.929", "fwd_time": "1.500", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 60, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000030428", "samples": "46025", "damp": "0.01000", "time": "8.953", "fwd_time": "1.500", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 60, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000003741", "samples": "46025", "damp": "0.01000", "time": "12.082", "fwd_time": "2.895", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 61, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000021188", "samples": "46025", "damp": "0.01000", "time": "3.260", "fwd_time": "0.515", "(v)ram": "cuda 10.24G" }, { "process": "gptq", "layer": 61, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000012185", "samples": "46025", "damp": "0.01000", "time": "3.083", "fwd_time": "0.480", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 61, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000004090", "samples": "46025", "damp": "0.01000", "time": "3.689", "fwd_time": "1.056", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 61, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000030536", "samples": "46025", "damp": "0.01000", "time": "9.118", "fwd_time": "1.482", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 61, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000032637", "samples": "46025", "damp": "0.01000", "time": "9.142", "fwd_time": "1.482", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 61, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000005517", "samples": "46025", "damp": "0.01000", "time": "11.295", "fwd_time": "2.846", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 62, "module": "linear_attn.in_proj_qkv", "feat: in, out": "5120, 10240", "dtype: size": "bf16: 103.1MB", "loss": "0.0000022684", "samples": "46025", "damp": "0.01000", "time": "3.467", "fwd_time": "0.551", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 62, "module": "linear_attn.in_proj_z", "feat: in, out": "5120, 6144", "dtype: size": "bf16: 61.9MB", "loss": "0.0000011953", "samples": "46025", "damp": "0.01000", "time": "3.295", "fwd_time": "0.481", "(v)ram": "cuda 10.24G" }, { "process": "gptq", "layer": 62, "module": "linear_attn.out_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000011525", "samples": "46025", "damp": "0.01000", "time": "3.548", "fwd_time": "1.070", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 62, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000030639", "samples": "46025", "damp": "0.01000", "time": "7.978", "fwd_time": "1.523", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 62, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000028734", "samples": "46025", "damp": "0.01000", "time": "8.066", "fwd_time": "1.523", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 62, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000009548", "samples": "46025", "damp": "0.01000", "time": "12.417", "fwd_time": "2.880", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 63, "module": "self_attn.v_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000008385", "samples": "46025", "damp": "0.01000", "time": "17.097", "fwd_time": "0.846", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 63, "module": "self_attn.q_proj", "feat: in, out": "5120, 12288", "dtype: size": "bf16: 123.8MB", "loss": "0.0000024606", "samples": "46025", "damp": "0.01000", "time": "17.212", "fwd_time": "0.846", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 63, "module": "self_attn.k_proj", "feat: in, out": "5120, 1024", "dtype: size": "bf16: 10.3MB", "loss": "0.0000002214", "samples": "46025", "damp": "0.01000", "time": "17.271", "fwd_time": "0.846", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 63, "module": "self_attn.o_proj", "feat: in, out": "6144, 5120", "dtype: size": "bf16: 61.9MB", "loss": "0.0000013553", "samples": "46025", "damp": "0.01000", "time": "3.893", "fwd_time": "0.914", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 63, "module": "mlp.up_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000026020", "samples": "46025", "damp": "0.01000", "time": "8.428", "fwd_time": "1.244", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 63, "module": "mlp.gate_proj", "feat: in, out": "5120, 17408", "dtype: size": "bf16: 175.3MB", "loss": "0.0000028705", "samples": "46025", "damp": "0.01000", "time": "8.466", "fwd_time": "1.244", "(v)ram": "cuda 10.25G" }, { "process": "gptq", "layer": 63, "module": "mlp.down_proj", "feat: in, out": "17408, 5120", "dtype: size": "bf16: 175.4MB", "loss": "0.0000021961", "samples": "46025", "damp": "0.01000", "time": "11.894", "fwd_time": "2.590", "(v)ram": "cuda 10.26G" } ] }