Add gt_mha_exact restoration manifest
Browse files
BERT-Base Checkpoints/gt_mha_exact/bert_base_gt_mha_exact_b4g8h12_random_seed42_explicit_fuseqkv_v2/checkpoint-100000/bert_gt_mha_manifest.json
ADDED
|
@@ -0,0 +1,368 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_name_or_path": "google-bert/bert-base-uncased",
|
| 3 |
+
"initialization": "random",
|
| 4 |
+
"attention_type": "gt_mha_exact",
|
| 5 |
+
"num_base_heads": 4,
|
| 6 |
+
"num_generators": 8,
|
| 7 |
+
"generator_mixing": "softmax",
|
| 8 |
+
"use_sdpa": false,
|
| 9 |
+
"fuse_base_qkv": true,
|
| 10 |
+
"sdpa_gqa_mode": "auto",
|
| 11 |
+
"enforce_paper_gt_mha": true,
|
| 12 |
+
"validation_mask_seed": 17029,
|
| 13 |
+
"validation_masking": "fixed_per_example_bert_80_10_10",
|
| 14 |
+
"head_coordinate_weight_decay": 0.0,
|
| 15 |
+
"head_coordinate_parameters": [
|
| 16 |
+
"bert.encoder.layer.0.attention.self.gt_attention.theta",
|
| 17 |
+
"bert.encoder.layer.0.attention.self.gt_attention.value_theta",
|
| 18 |
+
"bert.encoder.layer.1.attention.self.gt_attention.theta",
|
| 19 |
+
"bert.encoder.layer.1.attention.self.gt_attention.value_theta",
|
| 20 |
+
"bert.encoder.layer.10.attention.self.gt_attention.theta",
|
| 21 |
+
"bert.encoder.layer.10.attention.self.gt_attention.value_theta",
|
| 22 |
+
"bert.encoder.layer.11.attention.self.gt_attention.theta",
|
| 23 |
+
"bert.encoder.layer.11.attention.self.gt_attention.value_theta",
|
| 24 |
+
"bert.encoder.layer.2.attention.self.gt_attention.theta",
|
| 25 |
+
"bert.encoder.layer.2.attention.self.gt_attention.value_theta",
|
| 26 |
+
"bert.encoder.layer.3.attention.self.gt_attention.theta",
|
| 27 |
+
"bert.encoder.layer.3.attention.self.gt_attention.value_theta",
|
| 28 |
+
"bert.encoder.layer.4.attention.self.gt_attention.theta",
|
| 29 |
+
"bert.encoder.layer.4.attention.self.gt_attention.value_theta",
|
| 30 |
+
"bert.encoder.layer.5.attention.self.gt_attention.theta",
|
| 31 |
+
"bert.encoder.layer.5.attention.self.gt_attention.value_theta",
|
| 32 |
+
"bert.encoder.layer.6.attention.self.gt_attention.theta",
|
| 33 |
+
"bert.encoder.layer.6.attention.self.gt_attention.value_theta",
|
| 34 |
+
"bert.encoder.layer.7.attention.self.gt_attention.theta",
|
| 35 |
+
"bert.encoder.layer.7.attention.self.gt_attention.value_theta",
|
| 36 |
+
"bert.encoder.layer.8.attention.self.gt_attention.theta",
|
| 37 |
+
"bert.encoder.layer.8.attention.self.gt_attention.value_theta",
|
| 38 |
+
"bert.encoder.layer.9.attention.self.gt_attention.theta",
|
| 39 |
+
"bert.encoder.layer.9.attention.self.gt_attention.value_theta"
|
| 40 |
+
],
|
| 41 |
+
"teacher_student_distillation": false,
|
| 42 |
+
"replacement_audit": [
|
| 43 |
+
{
|
| 44 |
+
"layer": 0,
|
| 45 |
+
"attention_type": "gt_mha_exact",
|
| 46 |
+
"initialization": "random",
|
| 47 |
+
"teacher_attention_parameters": 1771776,
|
| 48 |
+
"gt_mha_attention_parameters": 656320,
|
| 49 |
+
"attention_parameter_reduction": 1115456,
|
| 50 |
+
"gt_mha_config": {
|
| 51 |
+
"hidden_size": 768,
|
| 52 |
+
"num_attention_heads": 12,
|
| 53 |
+
"attention_probs_dropout_prob": 0.1,
|
| 54 |
+
"initializer_range": 0.02,
|
| 55 |
+
"attention_type": "gt_mha_exact",
|
| 56 |
+
"num_base_heads": 4,
|
| 57 |
+
"num_generators": 8,
|
| 58 |
+
"generator_type": "full",
|
| 59 |
+
"generator_mixing": "softmax",
|
| 60 |
+
"use_sdpa": false,
|
| 61 |
+
"fuse_base_qkv": true,
|
| 62 |
+
"sdpa_gqa_mode": "auto",
|
| 63 |
+
"attention_bias": true
|
| 64 |
+
}
|
| 65 |
+
},
|
| 66 |
+
{
|
| 67 |
+
"layer": 1,
|
| 68 |
+
"attention_type": "gt_mha_exact",
|
| 69 |
+
"initialization": "random",
|
| 70 |
+
"teacher_attention_parameters": 1771776,
|
| 71 |
+
"gt_mha_attention_parameters": 656320,
|
| 72 |
+
"attention_parameter_reduction": 1115456,
|
| 73 |
+
"gt_mha_config": {
|
| 74 |
+
"hidden_size": 768,
|
| 75 |
+
"num_attention_heads": 12,
|
| 76 |
+
"attention_probs_dropout_prob": 0.1,
|
| 77 |
+
"initializer_range": 0.02,
|
| 78 |
+
"attention_type": "gt_mha_exact",
|
| 79 |
+
"num_base_heads": 4,
|
| 80 |
+
"num_generators": 8,
|
| 81 |
+
"generator_type": "full",
|
| 82 |
+
"generator_mixing": "softmax",
|
| 83 |
+
"use_sdpa": false,
|
| 84 |
+
"fuse_base_qkv": true,
|
| 85 |
+
"sdpa_gqa_mode": "auto",
|
| 86 |
+
"attention_bias": true
|
| 87 |
+
}
|
| 88 |
+
},
|
| 89 |
+
{
|
| 90 |
+
"layer": 2,
|
| 91 |
+
"attention_type": "gt_mha_exact",
|
| 92 |
+
"initialization": "random",
|
| 93 |
+
"teacher_attention_parameters": 1771776,
|
| 94 |
+
"gt_mha_attention_parameters": 656320,
|
| 95 |
+
"attention_parameter_reduction": 1115456,
|
| 96 |
+
"gt_mha_config": {
|
| 97 |
+
"hidden_size": 768,
|
| 98 |
+
"num_attention_heads": 12,
|
| 99 |
+
"attention_probs_dropout_prob": 0.1,
|
| 100 |
+
"initializer_range": 0.02,
|
| 101 |
+
"attention_type": "gt_mha_exact",
|
| 102 |
+
"num_base_heads": 4,
|
| 103 |
+
"num_generators": 8,
|
| 104 |
+
"generator_type": "full",
|
| 105 |
+
"generator_mixing": "softmax",
|
| 106 |
+
"use_sdpa": false,
|
| 107 |
+
"fuse_base_qkv": true,
|
| 108 |
+
"sdpa_gqa_mode": "auto",
|
| 109 |
+
"attention_bias": true
|
| 110 |
+
}
|
| 111 |
+
},
|
| 112 |
+
{
|
| 113 |
+
"layer": 3,
|
| 114 |
+
"attention_type": "gt_mha_exact",
|
| 115 |
+
"initialization": "random",
|
| 116 |
+
"teacher_attention_parameters": 1771776,
|
| 117 |
+
"gt_mha_attention_parameters": 656320,
|
| 118 |
+
"attention_parameter_reduction": 1115456,
|
| 119 |
+
"gt_mha_config": {
|
| 120 |
+
"hidden_size": 768,
|
| 121 |
+
"num_attention_heads": 12,
|
| 122 |
+
"attention_probs_dropout_prob": 0.1,
|
| 123 |
+
"initializer_range": 0.02,
|
| 124 |
+
"attention_type": "gt_mha_exact",
|
| 125 |
+
"num_base_heads": 4,
|
| 126 |
+
"num_generators": 8,
|
| 127 |
+
"generator_type": "full",
|
| 128 |
+
"generator_mixing": "softmax",
|
| 129 |
+
"use_sdpa": false,
|
| 130 |
+
"fuse_base_qkv": true,
|
| 131 |
+
"sdpa_gqa_mode": "auto",
|
| 132 |
+
"attention_bias": true
|
| 133 |
+
}
|
| 134 |
+
},
|
| 135 |
+
{
|
| 136 |
+
"layer": 4,
|
| 137 |
+
"attention_type": "gt_mha_exact",
|
| 138 |
+
"initialization": "random",
|
| 139 |
+
"teacher_attention_parameters": 1771776,
|
| 140 |
+
"gt_mha_attention_parameters": 656320,
|
| 141 |
+
"attention_parameter_reduction": 1115456,
|
| 142 |
+
"gt_mha_config": {
|
| 143 |
+
"hidden_size": 768,
|
| 144 |
+
"num_attention_heads": 12,
|
| 145 |
+
"attention_probs_dropout_prob": 0.1,
|
| 146 |
+
"initializer_range": 0.02,
|
| 147 |
+
"attention_type": "gt_mha_exact",
|
| 148 |
+
"num_base_heads": 4,
|
| 149 |
+
"num_generators": 8,
|
| 150 |
+
"generator_type": "full",
|
| 151 |
+
"generator_mixing": "softmax",
|
| 152 |
+
"use_sdpa": false,
|
| 153 |
+
"fuse_base_qkv": true,
|
| 154 |
+
"sdpa_gqa_mode": "auto",
|
| 155 |
+
"attention_bias": true
|
| 156 |
+
}
|
| 157 |
+
},
|
| 158 |
+
{
|
| 159 |
+
"layer": 5,
|
| 160 |
+
"attention_type": "gt_mha_exact",
|
| 161 |
+
"initialization": "random",
|
| 162 |
+
"teacher_attention_parameters": 1771776,
|
| 163 |
+
"gt_mha_attention_parameters": 656320,
|
| 164 |
+
"attention_parameter_reduction": 1115456,
|
| 165 |
+
"gt_mha_config": {
|
| 166 |
+
"hidden_size": 768,
|
| 167 |
+
"num_attention_heads": 12,
|
| 168 |
+
"attention_probs_dropout_prob": 0.1,
|
| 169 |
+
"initializer_range": 0.02,
|
| 170 |
+
"attention_type": "gt_mha_exact",
|
| 171 |
+
"num_base_heads": 4,
|
| 172 |
+
"num_generators": 8,
|
| 173 |
+
"generator_type": "full",
|
| 174 |
+
"generator_mixing": "softmax",
|
| 175 |
+
"use_sdpa": false,
|
| 176 |
+
"fuse_base_qkv": true,
|
| 177 |
+
"sdpa_gqa_mode": "auto",
|
| 178 |
+
"attention_bias": true
|
| 179 |
+
}
|
| 180 |
+
},
|
| 181 |
+
{
|
| 182 |
+
"layer": 6,
|
| 183 |
+
"attention_type": "gt_mha_exact",
|
| 184 |
+
"initialization": "random",
|
| 185 |
+
"teacher_attention_parameters": 1771776,
|
| 186 |
+
"gt_mha_attention_parameters": 656320,
|
| 187 |
+
"attention_parameter_reduction": 1115456,
|
| 188 |
+
"gt_mha_config": {
|
| 189 |
+
"hidden_size": 768,
|
| 190 |
+
"num_attention_heads": 12,
|
| 191 |
+
"attention_probs_dropout_prob": 0.1,
|
| 192 |
+
"initializer_range": 0.02,
|
| 193 |
+
"attention_type": "gt_mha_exact",
|
| 194 |
+
"num_base_heads": 4,
|
| 195 |
+
"num_generators": 8,
|
| 196 |
+
"generator_type": "full",
|
| 197 |
+
"generator_mixing": "softmax",
|
| 198 |
+
"use_sdpa": false,
|
| 199 |
+
"fuse_base_qkv": true,
|
| 200 |
+
"sdpa_gqa_mode": "auto",
|
| 201 |
+
"attention_bias": true
|
| 202 |
+
}
|
| 203 |
+
},
|
| 204 |
+
{
|
| 205 |
+
"layer": 7,
|
| 206 |
+
"attention_type": "gt_mha_exact",
|
| 207 |
+
"initialization": "random",
|
| 208 |
+
"teacher_attention_parameters": 1771776,
|
| 209 |
+
"gt_mha_attention_parameters": 656320,
|
| 210 |
+
"attention_parameter_reduction": 1115456,
|
| 211 |
+
"gt_mha_config": {
|
| 212 |
+
"hidden_size": 768,
|
| 213 |
+
"num_attention_heads": 12,
|
| 214 |
+
"attention_probs_dropout_prob": 0.1,
|
| 215 |
+
"initializer_range": 0.02,
|
| 216 |
+
"attention_type": "gt_mha_exact",
|
| 217 |
+
"num_base_heads": 4,
|
| 218 |
+
"num_generators": 8,
|
| 219 |
+
"generator_type": "full",
|
| 220 |
+
"generator_mixing": "softmax",
|
| 221 |
+
"use_sdpa": false,
|
| 222 |
+
"fuse_base_qkv": true,
|
| 223 |
+
"sdpa_gqa_mode": "auto",
|
| 224 |
+
"attention_bias": true
|
| 225 |
+
}
|
| 226 |
+
},
|
| 227 |
+
{
|
| 228 |
+
"layer": 8,
|
| 229 |
+
"attention_type": "gt_mha_exact",
|
| 230 |
+
"initialization": "random",
|
| 231 |
+
"teacher_attention_parameters": 1771776,
|
| 232 |
+
"gt_mha_attention_parameters": 656320,
|
| 233 |
+
"attention_parameter_reduction": 1115456,
|
| 234 |
+
"gt_mha_config": {
|
| 235 |
+
"hidden_size": 768,
|
| 236 |
+
"num_attention_heads": 12,
|
| 237 |
+
"attention_probs_dropout_prob": 0.1,
|
| 238 |
+
"initializer_range": 0.02,
|
| 239 |
+
"attention_type": "gt_mha_exact",
|
| 240 |
+
"num_base_heads": 4,
|
| 241 |
+
"num_generators": 8,
|
| 242 |
+
"generator_type": "full",
|
| 243 |
+
"generator_mixing": "softmax",
|
| 244 |
+
"use_sdpa": false,
|
| 245 |
+
"fuse_base_qkv": true,
|
| 246 |
+
"sdpa_gqa_mode": "auto",
|
| 247 |
+
"attention_bias": true
|
| 248 |
+
}
|
| 249 |
+
},
|
| 250 |
+
{
|
| 251 |
+
"layer": 9,
|
| 252 |
+
"attention_type": "gt_mha_exact",
|
| 253 |
+
"initialization": "random",
|
| 254 |
+
"teacher_attention_parameters": 1771776,
|
| 255 |
+
"gt_mha_attention_parameters": 656320,
|
| 256 |
+
"attention_parameter_reduction": 1115456,
|
| 257 |
+
"gt_mha_config": {
|
| 258 |
+
"hidden_size": 768,
|
| 259 |
+
"num_attention_heads": 12,
|
| 260 |
+
"attention_probs_dropout_prob": 0.1,
|
| 261 |
+
"initializer_range": 0.02,
|
| 262 |
+
"attention_type": "gt_mha_exact",
|
| 263 |
+
"num_base_heads": 4,
|
| 264 |
+
"num_generators": 8,
|
| 265 |
+
"generator_type": "full",
|
| 266 |
+
"generator_mixing": "softmax",
|
| 267 |
+
"use_sdpa": false,
|
| 268 |
+
"fuse_base_qkv": true,
|
| 269 |
+
"sdpa_gqa_mode": "auto",
|
| 270 |
+
"attention_bias": true
|
| 271 |
+
}
|
| 272 |
+
},
|
| 273 |
+
{
|
| 274 |
+
"layer": 10,
|
| 275 |
+
"attention_type": "gt_mha_exact",
|
| 276 |
+
"initialization": "random",
|
| 277 |
+
"teacher_attention_parameters": 1771776,
|
| 278 |
+
"gt_mha_attention_parameters": 656320,
|
| 279 |
+
"attention_parameter_reduction": 1115456,
|
| 280 |
+
"gt_mha_config": {
|
| 281 |
+
"hidden_size": 768,
|
| 282 |
+
"num_attention_heads": 12,
|
| 283 |
+
"attention_probs_dropout_prob": 0.1,
|
| 284 |
+
"initializer_range": 0.02,
|
| 285 |
+
"attention_type": "gt_mha_exact",
|
| 286 |
+
"num_base_heads": 4,
|
| 287 |
+
"num_generators": 8,
|
| 288 |
+
"generator_type": "full",
|
| 289 |
+
"generator_mixing": "softmax",
|
| 290 |
+
"use_sdpa": false,
|
| 291 |
+
"fuse_base_qkv": true,
|
| 292 |
+
"sdpa_gqa_mode": "auto",
|
| 293 |
+
"attention_bias": true
|
| 294 |
+
}
|
| 295 |
+
},
|
| 296 |
+
{
|
| 297 |
+
"layer": 11,
|
| 298 |
+
"attention_type": "gt_mha_exact",
|
| 299 |
+
"initialization": "random",
|
| 300 |
+
"teacher_attention_parameters": 1771776,
|
| 301 |
+
"gt_mha_attention_parameters": 656320,
|
| 302 |
+
"attention_parameter_reduction": 1115456,
|
| 303 |
+
"gt_mha_config": {
|
| 304 |
+
"hidden_size": 768,
|
| 305 |
+
"num_attention_heads": 12,
|
| 306 |
+
"attention_probs_dropout_prob": 0.1,
|
| 307 |
+
"initializer_range": 0.02,
|
| 308 |
+
"attention_type": "gt_mha_exact",
|
| 309 |
+
"num_base_heads": 4,
|
| 310 |
+
"num_generators": 8,
|
| 311 |
+
"generator_type": "full",
|
| 312 |
+
"generator_mixing": "softmax",
|
| 313 |
+
"use_sdpa": false,
|
| 314 |
+
"fuse_base_qkv": true,
|
| 315 |
+
"sdpa_gqa_mode": "auto",
|
| 316 |
+
"attention_bias": true
|
| 317 |
+
}
|
| 318 |
+
}
|
| 319 |
+
],
|
| 320 |
+
"parameter_counts": {
|
| 321 |
+
"model_parameter_count": 96128826,
|
| 322 |
+
"trainable_parameter_count": 96128826,
|
| 323 |
+
"self_attention_parameter_count": 7875840,
|
| 324 |
+
"attention_output_parameter_count": 7105536,
|
| 325 |
+
"attention_block_parameter_count": 14981376
|
| 326 |
+
},
|
| 327 |
+
"arguments": {
|
| 328 |
+
"model_name_or_path": "google-bert/bert-base-uncased",
|
| 329 |
+
"attention_type": "gt_mha_exact",
|
| 330 |
+
"initialization": "random",
|
| 331 |
+
"num_base_heads": 4,
|
| 332 |
+
"num_generators": 8,
|
| 333 |
+
"generator_mixing": "softmax",
|
| 334 |
+
"use_sdpa": false,
|
| 335 |
+
"fuse_base_qkv": true,
|
| 336 |
+
"sdpa_gqa_mode": "auto",
|
| 337 |
+
"enforce_paper_gt_mha": true,
|
| 338 |
+
"dataset_name": "/work/hdd/bifg/arai3/lgma_data/wikitext-103-raw-v1",
|
| 339 |
+
"dataset_config": "",
|
| 340 |
+
"train_split": "train",
|
| 341 |
+
"validation_split": "validation",
|
| 342 |
+
"text_column": "text",
|
| 343 |
+
"max_sequence_length": 128,
|
| 344 |
+
"mlm_probability": 0.15,
|
| 345 |
+
"validation_mask_seed": 17029,
|
| 346 |
+
"output_dir": "/work/hdd/bifg/arai3/lgma_runs/bert_base_gt_mha_exact_b4g8h12_random_seed42_explicit_fuseqkv_v2",
|
| 347 |
+
"per_device_train_batch_size": 32,
|
| 348 |
+
"per_device_eval_batch_size": 32,
|
| 349 |
+
"gradient_accumulation_steps": 2,
|
| 350 |
+
"learning_rate": 5e-05,
|
| 351 |
+
"weight_decay": 0.01,
|
| 352 |
+
"warmup_steps": 10000,
|
| 353 |
+
"max_steps": 100000,
|
| 354 |
+
"logging_steps": 100,
|
| 355 |
+
"eval_steps": 1000,
|
| 356 |
+
"save_steps": 1000,
|
| 357 |
+
"save_total_limit": 3,
|
| 358 |
+
"seed": 42,
|
| 359 |
+
"num_workers": 4,
|
| 360 |
+
"bf16": true,
|
| 361 |
+
"fp16": false,
|
| 362 |
+
"gradient_checkpointing": true,
|
| 363 |
+
"resume_from_checkpoint": null,
|
| 364 |
+
"report_to": "wandb",
|
| 365 |
+
"run_name": "bert_base_gt_mha_exact_b4g8h12_random_seed42_explicit_fuseqkv_v2",
|
| 366 |
+
"trust_remote_code": false
|
| 367 |
+
}
|
| 368 |
+
}
|