Update config.json
Browse files- config.json +52 -32
config.json
CHANGED
|
@@ -3,42 +3,62 @@
|
|
| 3 |
"architectures": [
|
| 4 |
"SuperLinearForCausalLM"
|
| 5 |
],
|
| 6 |
-
"auto_map": {
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
},
|
| 10 |
-
|
| 11 |
-
"
|
| 12 |
-
"
|
| 13 |
-
"
|
| 14 |
-
"
|
| 15 |
-
"
|
| 16 |
"inf_pred_len": 96,
|
| 17 |
-
"layer_type": "RLinear",
|
| 18 |
-
"linear_checkpoints_dir": "checkpoints5",
|
| 19 |
-
"linear_checkpoints_path": "/cs/azencot_fsas/MoE/",
|
| 20 |
-
"load_linear": 0,
|
| 21 |
-
"load_weights" :0,
|
| 22 |
"max_horizon": 96,
|
| 23 |
-
"
|
| 24 |
-
|
|
|
|
| 25 |
"moe": 1,
|
| 26 |
-
"moe_n_experts":
|
| 27 |
-
"
|
| 28 |
"noisy_gating_std": 0.1,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 29 |
"noisy_gating_std_decay": 1,
|
| 30 |
-
"
|
| 31 |
-
"
|
| 32 |
-
"
|
| 33 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
| 34 |
"torch_dtype": "float32",
|
| 35 |
-
"transformers_version": "4.40.1"
|
| 36 |
-
"use_fft": 1,
|
| 37 |
-
"train_epochs": 30,
|
| 38 |
-
"patience": 5,
|
| 39 |
-
"lradj": "type1",
|
| 40 |
-
"learning_rate": 0.005,
|
| 41 |
-
"channel_ind": 0,
|
| 42 |
-
"misc_moe" :12,
|
| 43 |
-
"full_size":0
|
| 44 |
}
|
|
|
|
| 3 |
"architectures": [
|
| 4 |
"SuperLinearForCausalLM"
|
| 5 |
],
|
| 6 |
+
"auto_map": {
|
| 7 |
+
"AutoConfig": "configuration_super_linear.SuperLinearConfig",
|
| 8 |
+
"AutoModelForCausalLM": "modeling_super_linear.SuperLinearForCausalLM"
|
| 9 |
+
},
|
| 10 |
+
|
| 11 |
+
"_comment_model_architecture": "Model architecture parameters",
|
| 12 |
+
"train_seq_len": 512,
|
| 13 |
+
"train_pred_len": 96,
|
| 14 |
+
"seq_len": 512,
|
| 15 |
+
"pred_len": 96,
|
| 16 |
"inf_pred_len": 96,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
"max_horizon": 96,
|
| 18 |
+
"auto_regressive": 1,
|
| 19 |
+
|
| 20 |
+
"_comment_moe": "MoE (Mixture of Experts) parameters",
|
| 21 |
"moe": 1,
|
| 22 |
+
"moe_n_experts": 4,
|
| 23 |
+
"top_k_experts": 12,
|
| 24 |
"noisy_gating_std": 0.1,
|
| 25 |
+
"moe_temp": 1.0,
|
| 26 |
+
"moe_norm": false,
|
| 27 |
+
"layer_type": "RLinear",
|
| 28 |
+
"n_experts": 4,
|
| 29 |
+
"comp_moe": 12,
|
| 30 |
+
"freeze_experts": true,
|
| 31 |
+
|
| 32 |
+
"_comment_fft": "FFT-based gating parameters",
|
| 33 |
+
"use_fft": true,
|
| 34 |
+
"fft_len": 5000,
|
| 35 |
+
|
| 36 |
+
"_comment_experts": "Expert configuration",
|
| 37 |
+
"freq_experts": "mean_naive_1/4_1/6_1/7_1/8_1/12_1/14_1/16_1/21_1/24_1/28_1/30_1/32_1/36_1/42_1/48_1/52_1/56_1/60_1/72_1/84_1/90_1/96_1/120_1/144_1/168_1/180_1/224_1/252_1/288_1/336_1/365_1/504_1/672_1/1008_1/1440_1/2016_1/3600",
|
| 38 |
+
|
| 39 |
+
"_comment_loading": "Model loading and saving parameters",
|
| 40 |
+
"load_linear": true,
|
| 41 |
+
"load_weights_full": true,
|
| 42 |
+
"linear_freq_weights_path": "./weights/linear_freq_weights/",
|
| 43 |
+
"full_weights_path": "./weights/full_weights/checkpoint.pth",
|
| 44 |
+
|
| 45 |
+
"_comment_training": "Training parameters",
|
| 46 |
+
"resample_long_lookback": false,
|
| 47 |
+
|
| 48 |
+
"_comment_legacy": "Legacy parameters for backward compatibility",
|
| 49 |
+
"linear_checkpoints_path": "/cs/azencot_fsas/MoE/",
|
| 50 |
+
"linear_checkpoints_dir": "checkpoints5",
|
| 51 |
+
"manual_moe": 0,
|
| 52 |
+
"misc_moe": 1,
|
| 53 |
"noisy_gating_std_decay": 1,
|
| 54 |
+
"ker_len": 50,
|
| 55 |
+
"con": 0,
|
| 56 |
+
"d_model": 512,
|
| 57 |
+
"mlp_gating": 1,
|
| 58 |
+
"dropout": 0.0,
|
| 59 |
+
|
| 60 |
+
"_comment_system": "System and framework parameters",
|
| 61 |
+
"model_type": "super_linear",
|
| 62 |
"torch_dtype": "float32",
|
| 63 |
+
"transformers_version": "4.40.1"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 64 |
}
|