remove ir

Neural_Engine_INT8_IR/conf.yaml

@@ -1,2299 +0,0 @@
-model:
-  name: model
-  operator:
-    input_data:
-      type: Input
-      output:
-        input_ids:0:
-          dtype: int32
-          shape: [-1, -1]
-        segment_ids:0:
-          dtype: int32
-          shape: [-1, -1]
-        input_mask:0:
-          dtype: int32
-          shape: [-1, -1]
-        bert.embeddings.position_embeddings.weight:0:
-          dtype: fp32
-          shape: [512, 256]
-          location: [0, 524288]
-        bert.embeddings.word_embeddings.weight:0:
-          dtype: fp32
-          shape: [30522, 256]
-          location: [524288, 31254528]
-        bert.embeddings.token_type_embeddings.weight:0:
-          dtype: fp32
-          shape: [2, 256]
-          location: [31778816, 2048]
-        bert.embeddings.LayerNorm.weight:0:
-          dtype: fp32
-          shape: [256]
-          location: [31780864, 1024]
-        bert.embeddings.LayerNorm.bias:0:
-          dtype: fp32
-          shape: [256]
-          location: [31781888, 1024]
-        111:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [31782912, 4]
-        111:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [31782916, 4]
-        '576:0':
-          dtype: s8
-          shape: [256, 256]
-          location: [31782920, 65536]
-        bert.encoder.layer.0.attention.self.key.bias:0:
-          dtype: s32
-          shape: [256]
-          location: [31848456, 1024]
-        111:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [31988776, 4]
-        111:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [31988780, 4]
-        576:0_min:
-          dtype: fp32
-          shape: [256]
-          location: [31849480, 1024]
-        576:0_max:
-          dtype: fp32
-          shape: [256]
-          location: [31850504, 1024]
-        Add_34:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [31851536, 4]
-        Add_34:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [31851540, 4]
-        '579:0':
-          dtype: s8
-          shape: [256, 256]
-          location: [31851544, 65536]
-        bert.encoder.layer.0.attention.self.value.bias:0:
-          dtype: s32
-          shape: [256]
-          location: [31917080, 1024]
-        579:0_min:
-          dtype: fp32
-          shape: [256]
-          location: [31918104, 1024]
-        579:0_max:
-          dtype: fp32
-          shape: [256]
-          location: [31919128, 1024]
-        Add_46:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [31920160, 4]
-        Add_46:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [31920164, 4]
-        '575:0':
-          dtype: s8
-          shape: [256, 256]
-          location: [31920168, 65536]
-        bert.encoder.layer.0.attention.self.query.bias:0:
-          dtype: s32
-          shape: [256]
-          location: [31985704, 1024]
-        575:0_min:
-          dtype: fp32
-          shape: [256]
-          location: [31986728, 1024]
-        575:0_max:
-          dtype: fp32
-          shape: [256]
-          location: [31987752, 1024]
-        Add_32:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [31988784, 4]
-        Add_32:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [31988788, 4]
-        163:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [31988792, 4]
-        163:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [31988796, 4]
-        131:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [31988800, 4]
-        131:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [31988804, 4]
-        169:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [31988808, 4]
-        169:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [31988812, 4]
-        170:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [31988824, 4]
-        170:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [31988828, 4]
-        148:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [31988832, 4]
-        148:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [31988836, 4]
-        172:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [31988840, 4]
-        172:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [31988844, 4]
-        '585:0':
-          dtype: s8
-          shape: [256, 256]
-          location: [31988848, 65536]
-        bert.encoder.layer.0.attention.output.dense.bias:0:
-          dtype: s32
-          shape: [256]
-          location: [32054384, 1024]
-        184:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [32057456, 4]
-        184:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [32057460, 4]
-        585:0_min:
-          dtype: fp32
-          shape: [256]
-          location: [32055408, 1024]
-        585:0_max:
-          dtype: fp32
-          shape: [256]
-          location: [32056432, 1024]
-        188:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [32057464, 4]
-        188:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [32057468, 4]
-        bert.encoder.layer.0.attention.output.LayerNorm.weight:0:
-          dtype: fp32
-          shape: [256]
-          location: [32057472, 1024]
-        bert.encoder.layer.0.attention.output.LayerNorm.bias:0:
-          dtype: fp32
-          shape: [256]
-          location: [32058496, 1024]
-        199:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [32059520, 4]
-        199:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [32059524, 4]
-        '586:0':
-          dtype: s8
-          shape: [1024, 256]
-          location: [32059528, 262144]
-        bert.encoder.layer.0.intermediate.dense.bias:0:
-          dtype: s32
-          shape: [1024]
-          location: [32321672, 4096]
-        199:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [32333960, 4]
-        199:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [32333964, 4]
-        586:0_min:
-          dtype: fp32
-          shape: [1024]
-          location: [32325768, 4096]
-        586:0_max:
-          dtype: fp32
-          shape: [1024]
-          location: [32329864, 4096]
-        210:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [32599200, 4]
-        210:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [32599204, 4]
-        '587:0':
-          dtype: s8
-          shape: [256, 1024]
-          location: [32333984, 262144]
-        bert.encoder.layer.0.output.dense.bias:0:
-          dtype: s32
-          shape: [256]
-          location: [32596128, 1024]
-        587:0_min:
-          dtype: fp32
-          shape: [256]
-          location: [32597152, 1024]
-        587:0_max:
-          dtype: fp32
-          shape: [256]
-          location: [32598176, 1024]
-        214:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [32599208, 4]
-        214:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [32599212, 4]
-        bert.encoder.layer.0.output.LayerNorm.weight:0:
-          dtype: fp32
-          shape: [256]
-          location: [32599216, 1024]
-        bert.encoder.layer.0.output.LayerNorm.bias:0:
-          dtype: fp32
-          shape: [256]
-          location: [32600240, 1024]
-        225:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [32601264, 4]
-        225:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [32601268, 4]
-        '589:0':
-          dtype: s8
-          shape: [256, 256]
-          location: [32601272, 65536]
-        bert.encoder.layer.1.attention.self.key.bias:0:
-          dtype: s32
-          shape: [256]
-          location: [32666808, 1024]
-        225:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [32807128, 4]
-        225:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [32807132, 4]
-        589:0_min:
-          dtype: fp32
-          shape: [256]
-          location: [32667832, 1024]
-        589:0_max:
-          dtype: fp32
-          shape: [256]
-          location: [32668856, 1024]
-        Add_128:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [32669888, 4]
-        Add_128:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [32669892, 4]
-        '592:0':
-          dtype: s8
-          shape: [256, 256]
-          location: [32669896, 65536]
-        bert.encoder.layer.1.attention.self.value.bias:0:
-          dtype: s32
-          shape: [256]
-          location: [32735432, 1024]
-        592:0_min:
-          dtype: fp32
-          shape: [256]
-          location: [32736456, 1024]
-        592:0_max:
-          dtype: fp32
-          shape: [256]
-          location: [32737480, 1024]
-        Add_140:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [32738512, 4]
-        Add_140:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [32738516, 4]
-        '588:0':
-          dtype: s8
-          shape: [256, 256]
-          location: [32738520, 65536]
-        bert.encoder.layer.1.attention.self.query.bias:0:
-          dtype: s32
-          shape: [256]
-          location: [32804056, 1024]
-        588:0_min:
-          dtype: fp32
-          shape: [256]
-          location: [32805080, 1024]
-        588:0_max:
-          dtype: fp32
-          shape: [256]
-          location: [32806104, 1024]
-        Add_126:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [32807136, 4]
-        Add_126:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [32807140, 4]
-        277:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [32807144, 4]
-        277:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [32807148, 4]
-        245:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [32807152, 4]
-        245:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [32807156, 4]
-        283:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [32807160, 4]
-        283:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [32807164, 4]
-        284:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [32807176, 4]
-        284:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [32807180, 4]
-        262:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [32807184, 4]
-        262:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [32807188, 4]
-        286:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [32807192, 4]
-        286:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [32807196, 4]
-        '598:0':
-          dtype: s8
-          shape: [256, 256]
-          location: [32807200, 65536]
-        bert.encoder.layer.1.attention.output.dense.bias:0:
-          dtype: s32
-          shape: [256]
-          location: [32872736, 1024]
-        298:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [32875808, 4]
-        298:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [32875812, 4]
-        598:0_min:
-          dtype: fp32
-          shape: [256]
-          location: [32873760, 1024]
-        598:0_max:
-          dtype: fp32
-          shape: [256]
-          location: [32874784, 1024]
-        302:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [32875816, 4]
-        302:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [32875820, 4]
-        bert.encoder.layer.1.attention.output.LayerNorm.weight:0:
-          dtype: fp32
-          shape: [256]
-          location: [32875824, 1024]
-        bert.encoder.layer.1.attention.output.LayerNorm.bias:0:
-          dtype: fp32
-          shape: [256]
-          location: [32876848, 1024]
-        313:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [32877872, 4]
-        313:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [32877876, 4]
-        '599:0':
-          dtype: s8
-          shape: [1024, 256]
-          location: [32877880, 262144]
-        bert.encoder.layer.1.intermediate.dense.bias:0:
-          dtype: s32
-          shape: [1024]
-          location: [33140024, 4096]
-        313:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [33152312, 4]
-        313:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [33152316, 4]
-        599:0_min:
-          dtype: fp32
-          shape: [1024]
-          location: [33144120, 4096]
-        599:0_max:
-          dtype: fp32
-          shape: [1024]
-          location: [33148216, 4096]
-        324:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [33417552, 4]
-        324:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [33417556, 4]
-        '600:0':
-          dtype: s8
-          shape: [256, 1024]
-          location: [33152336, 262144]
-        bert.encoder.layer.1.output.dense.bias:0:
-          dtype: s32
-          shape: [256]
-          location: [33414480, 1024]
-        600:0_min:
-          dtype: fp32
-          shape: [256]
-          location: [33415504, 1024]
-        600:0_max:
-          dtype: fp32
-          shape: [256]
-          location: [33416528, 1024]
-        328:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [33417560, 4]
-        328:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [33417564, 4]
-        bert.encoder.layer.1.output.LayerNorm.weight:0:
-          dtype: fp32
-          shape: [256]
-          location: [33417568, 1024]
-        bert.encoder.layer.1.output.LayerNorm.bias:0:
-          dtype: fp32
-          shape: [256]
-          location: [33418592, 1024]
-        339:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [33419616, 4]
-        339:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [33419620, 4]
-        '602:0':
-          dtype: s8
-          shape: [256, 256]
-          location: [33419624, 65536]
-        bert.encoder.layer.2.attention.self.key.bias:0:
-          dtype: s32
-          shape: [256]
-          location: [33485160, 1024]
-        339:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [33625480, 4]
-        339:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [33625484, 4]
-        602:0_min:
-          dtype: fp32
-          shape: [256]
-          location: [33486184, 1024]
-        602:0_max:
-          dtype: fp32
-          shape: [256]
-          location: [33487208, 1024]
-        Add_222:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [33488240, 4]
-        Add_222:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [33488244, 4]
-        '605:0':
-          dtype: s8
-          shape: [256, 256]
-          location: [33488248, 65536]
-        bert.encoder.layer.2.attention.self.value.bias:0:
-          dtype: s32
-          shape: [256]
-          location: [33553784, 1024]
-        605:0_min:
-          dtype: fp32
-          shape: [256]
-          location: [33554808, 1024]
-        605:0_max:
-          dtype: fp32
-          shape: [256]
-          location: [33555832, 1024]
-        Add_234:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [33556864, 4]
-        Add_234:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [33556868, 4]
-        '601:0':
-          dtype: s8
-          shape: [256, 256]
-          location: [33556872, 65536]
-        bert.encoder.layer.2.attention.self.query.bias:0:
-          dtype: s32
-          shape: [256]
-          location: [33622408, 1024]
-        601:0_min:
-          dtype: fp32
-          shape: [256]
-          location: [33623432, 1024]
-        601:0_max:
-          dtype: fp32
-          shape: [256]
-          location: [33624456, 1024]
-        Add_220:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [33625488, 4]
-        Add_220:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [33625492, 4]
-        391:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [33625496, 4]
-        391:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [33625500, 4]
-        359:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [33625504, 4]
-        359:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [33625508, 4]
-        397:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [33625512, 4]
-        397:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [33625516, 4]
-        398:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [33625528, 4]
-        398:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [33625532, 4]
-        376:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [33625536, 4]
-        376:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [33625540, 4]
-        400:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [33625544, 4]
-        400:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [33625548, 4]
-        '611:0':
-          dtype: s8
-          shape: [256, 256]
-          location: [33625552, 65536]
-        bert.encoder.layer.2.attention.output.dense.bias:0:
-          dtype: s32
-          shape: [256]
-          location: [33691088, 1024]
-        412:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [33694160, 4]
-        412:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [33694164, 4]
-        611:0_min:
-          dtype: fp32
-          shape: [256]
-          location: [33692112, 1024]
-        611:0_max:
-          dtype: fp32
-          shape: [256]
-          location: [33693136, 1024]
-        416:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [33694168, 4]
-        416:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [33694172, 4]
-        bert.encoder.layer.2.attention.output.LayerNorm.weight:0:
-          dtype: fp32
-          shape: [256]
-          location: [33694176, 1024]
-        bert.encoder.layer.2.attention.output.LayerNorm.bias:0:
-          dtype: fp32
-          shape: [256]
-          location: [33695200, 1024]
-        427:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [33696224, 4]
-        427:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [33696228, 4]
-        '612:0':
-          dtype: s8
-          shape: [1024, 256]
-          location: [33696232, 262144]
-        bert.encoder.layer.2.intermediate.dense.bias:0:
-          dtype: s32
-          shape: [1024]
-          location: [33958376, 4096]
-        427:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [33970664, 4]
-        427:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [33970668, 4]
-        612:0_min:
-          dtype: fp32
-          shape: [1024]
-          location: [33962472, 4096]
-        612:0_max:
-          dtype: fp32
-          shape: [1024]
-          location: [33966568, 4096]
-        438:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [34235904, 4]
-        438:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [34235908, 4]
-        '613:0':
-          dtype: s8
-          shape: [256, 1024]
-          location: [33970688, 262144]
-        bert.encoder.layer.2.output.dense.bias:0:
-          dtype: s32
-          shape: [256]
-          location: [34232832, 1024]
-        613:0_min:
-          dtype: fp32
-          shape: [256]
-          location: [34233856, 1024]
-        613:0_max:
-          dtype: fp32
-          shape: [256]
-          location: [34234880, 1024]
-        442:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [34235912, 4]
-        442:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [34235916, 4]
-        bert.encoder.layer.2.output.LayerNorm.weight:0:
-          dtype: fp32
-          shape: [256]
-          location: [34235920, 1024]
-        bert.encoder.layer.2.output.LayerNorm.bias:0:
-          dtype: fp32
-          shape: [256]
-          location: [34236944, 1024]
-        453:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [34237968, 4]
-        453:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [34237972, 4]
-        '615:0':
-          dtype: s8
-          shape: [256, 256]
-          location: [34237976, 65536]
-        bert.encoder.layer.3.attention.self.key.bias:0:
-          dtype: s32
-          shape: [256]
-          location: [34303512, 1024]
-        453:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [34443832, 4]
-        453:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [34443836, 4]
-        615:0_min:
-          dtype: fp32
-          shape: [256]
-          location: [34304536, 1024]
-        615:0_max:
-          dtype: fp32
-          shape: [256]
-          location: [34305560, 1024]
-        Add_316:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [34306592, 4]
-        Add_316:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [34306596, 4]
-        '618:0':
-          dtype: s8
-          shape: [256, 256]
-          location: [34306600, 65536]
-        bert.encoder.layer.3.attention.self.value.bias:0:
-          dtype: s32
-          shape: [256]
-          location: [34372136, 1024]
-        618:0_min:
-          dtype: fp32
-          shape: [256]
-          location: [34373160, 1024]
-        618:0_max:
-          dtype: fp32
-          shape: [256]
-          location: [34374184, 1024]
-        Add_328:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [34375216, 4]
-        Add_328:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [34375220, 4]
-        '614:0':
-          dtype: s8
-          shape: [256, 256]
-          location: [34375224, 65536]
-        bert.encoder.layer.3.attention.self.query.bias:0:
-          dtype: s32
-          shape: [256]
-          location: [34440760, 1024]
-        614:0_min:
-          dtype: fp32
-          shape: [256]
-          location: [34441784, 1024]
-        614:0_max:
-          dtype: fp32
-          shape: [256]
-          location: [34442808, 1024]
-        Add_314:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [34443840, 4]
-        Add_314:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [34443844, 4]
-        505:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [34443848, 4]
-        505:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [34443852, 4]
-        473:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [34443856, 4]
-        473:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [34443860, 4]
-        511:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [34443864, 4]
-        511:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [34443868, 4]
-        512:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [34443880, 4]
-        512:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [34443884, 4]
-        490:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [34443888, 4]
-        490:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [34443892, 4]
-        514:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [34443896, 4]
-        514:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [34443900, 4]
-        '624:0':
-          dtype: s8
-          shape: [256, 256]
-          location: [34443904, 65536]
-        bert.encoder.layer.3.attention.output.dense.bias:0:
-          dtype: s32
-          shape: [256]
-          location: [34509440, 1024]
-        526:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [34512512, 4]
-        526:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [34512516, 4]
-        624:0_min:
-          dtype: fp32
-          shape: [256]
-          location: [34510464, 1024]
-        624:0_max:
-          dtype: fp32
-          shape: [256]
-          location: [34511488, 1024]
-        530:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [34512520, 4]
-        530:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [34512524, 4]
-        bert.encoder.layer.3.attention.output.LayerNorm.weight:0:
-          dtype: fp32
-          shape: [256]
-          location: [34512528, 1024]
-        bert.encoder.layer.3.attention.output.LayerNorm.bias:0:
-          dtype: fp32
-          shape: [256]
-          location: [34513552, 1024]
-        541:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [34514576, 4]
-        541:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [34514580, 4]
-        '625:0':
-          dtype: s8
-          shape: [1024, 256]
-          location: [34514584, 262144]
-        bert.encoder.layer.3.intermediate.dense.bias:0:
-          dtype: s32
-          shape: [1024]
-          location: [34776728, 4096]
-        541:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [34789016, 4]
-        541:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [34789020, 4]
-        625:0_min:
-          dtype: fp32
-          shape: [1024]
-          location: [34780824, 4096]
-        625:0_max:
-          dtype: fp32
-          shape: [1024]
-          location: [34784920, 4096]
-        552:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [35054256, 4]
-        552:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [35054260, 4]
-        '626:0':
-          dtype: s8
-          shape: [256, 1024]
-          location: [34789040, 262144]
-        bert.encoder.layer.3.output.dense.bias:0:
-          dtype: s32
-          shape: [256]
-          location: [35051184, 1024]
-        626:0_min:
-          dtype: fp32
-          shape: [256]
-          location: [35052208, 1024]
-        626:0_max:
-          dtype: fp32
-          shape: [256]
-          location: [35053232, 1024]
-        556:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [35054264, 4]
-        556:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [35054268, 4]
-        bert.encoder.layer.3.output.LayerNorm.weight:0:
-          dtype: fp32
-          shape: [256]
-          location: [35054272, 1024]
-        bert.encoder.layer.3.output.LayerNorm.bias:0:
-          dtype: fp32
-          shape: [256]
-          location: [35055296, 1024]
-        569:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [35056320, 4]
-        569:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [35056324, 4]
-        bert.pooler.dense.weight:0:
-          dtype: s8
-          shape: [256, 256]
-          location: [35056328, 65536]
-        bert.pooler.dense.bias:0:
-          dtype: s32
-          shape: [256]
-          location: [35121864, 1024]
-        569:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [35122888, 4]
-        569:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [35122892, 4]
-        bert.pooler.dense.weight:0_min:
-          dtype: fp32
-          shape: [256]
-          location: [35122896, 1024]
-        bert.pooler.dense.weight:0_max:
-          dtype: fp32
-          shape: [256]
-          location: [35123920, 1024]
-        571:0_quant_min:
-          dtype: fp32
-          shape: [1]
-          location: [35125472, 4]
-        571:0_quant_max:
-          dtype: fp32
-          shape: [1]
-          location: [35125476, 4]
-        classifier.weight:0:
-          dtype: s8
-          shape: [2, 256]
-          location: [35124952, 512]
-        classifier.bias:0:
-          dtype: s32
-          shape: [2]
-          location: [35125464, 8]
-        classifier.weight:0_min:
-          dtype: fp32
-          shape: [2]
-          location: [35125480, 8]
-        classifier.weight:0_max:
-          dtype: fp32
-          shape: [2]
-          location: [35125488, 8]
-        output:0_min:
-          dtype: fp32
-          shape: [1]
-          location: [35125496, 4]
-        output:0_max:
-          dtype: fp32
-          shape: [1]
-          location: [35125500, 4]
-    padding_sequence:
-      type: PaddingSequence
-      input:
-        input_mask:0: {}
-      output:
-        padding_sequence:0: {}
-      attr:
-        dst_shape: -1,4,0,-1
-        dims: 1
-    position_embeddings/after/reshape:
-      type: Reshape
-      input:
-        bert.embeddings.position_embeddings.weight:0: {}
-        input_ids:0: {}
-      output:
-        position_embeddings/after/reshape:0: {}
-      attr:
-        dst_shape: 1,-1,256
-        dims: 1
-    Gather_18:
-      type: Reshape
-      input:
-        position_embeddings/after/reshape:0: {}
-      output:
-        '99:0': {}
-      attr:
-        dst_shape: 1,-1
-    word_embeddings/reshape:
-      type: Reshape
-      input:
-        input_ids:0: {}
-      output:
-        word_embeddings/reshape:0: {}
-      attr:
-        dst_shape: -1
-    Gather_15:
-      type: Gather
-      input:
-        word_embeddings/reshape:0: {}
-        bert.embeddings.word_embeddings.weight:0: {}
-      output:
-        Gather_15:0: {}
-      attr:
-        axis: 0
-        batch_dims: 0
-    word_embeddings/after/reshape:
-      type: Reshape
-      input:
-        Gather_15:0: {}
-        input_ids:0: {}
-      output:
-        word_embeddings/after/reshape:0: {}
-      attr:
-        dst_shape: -1,-1,256
-        dims: 0,1
-    word_embeddings/add_reshape:
-      type: Reshape
-      input:
-        word_embeddings/after/reshape:0: {}
-        input_ids:0: {}
-      output:
-        word_embeddings/add_reshape:0: {}
-      attr:
-        dst_shape: -1,-1,256
-        dims: 0,1
-        mul: 1,2
-    token_type_embeddings/reshape:
-      type: Reshape
-      input:
-        segment_ids:0: {}
-      output:
-        token_type_embeddings/reshape:0: {}
-      attr:
-        dst_shape: -1
-    Gather_16:
-      type: Gather
-      input:
-        token_type_embeddings/reshape:0: {}
-        bert.embeddings.token_type_embeddings.weight:0: {}
-      output:
-        Gather_16:0: {}
-      attr:
-        axis: 0
-        batch_dims: 0
-    token_type_embeddings/after/reshape:
-      type: Reshape
-      input:
-        Gather_16:0: {}
-        segment_ids:0: {}
-      output:
-        token_type_embeddings/after/reshape:0: {}
-      attr:
-        dst_shape: -1,-1,256
-        dims: 0,1
-    token_type_embeddings/add_reshape:
-      type: Reshape
-      input:
-        token_type_embeddings/after/reshape:0: {}
-        segment_ids:0: {}
-      output:
-        token_type_embeddings/add_reshape:0: {}
-      attr:
-        dst_shape: -1,-1,256
-        dims: 0,1
-        mul: 1,2
-    Add_17:
-      type: BinaryAdd
-      input:
-        token_type_embeddings/add_reshape:0: {}
-        '99:0': {}
-        word_embeddings/add_reshape:0: {}
-      output:
-        Add_17:0: {}
-      attr:
-        append_op: sum
-    embeddings/after_add_reshape:
-      type: Reshape
-      input:
-        Add_17:0: {}
-        input_ids:0: {}
-      output:
-        embeddings/after_add_reshape:0: {}
-      attr:
-        dst_shape: -1,-1,256
-        dims: 0,1
-    embeddings_add/reshape_2d:
-      type: Reshape
-      input:
-        embeddings/after_add_reshape:0: {}
-      output:
-        embeddings_add/reshape_2d:0: {}
-      attr:
-        dst_shape: -1,256
-    Add_30:
-      type: LayerNorm
-      input:
-        embeddings_add/reshape_2d:0: {}
-        bert.embeddings.LayerNorm.weight:0: {}
-        bert.embeddings.LayerNorm.bias:0: {}
-      output:
-        '111:0': {}
-      attr:
-        epsilon: 9.999999960041972e-13
-    Add_30_reorder_post:
-      type: Reorder
-      input:
-        '111:0': {}
-      output:
-        111:0_reorder: {}
-      attr:
-        src_perm: 0,1
-        dst_perm: 1,0
-    Add_34_quant_0:
-      type: Quantize
-      input:
-        111:0_reorder: {}
-        111:0_min: {}
-        111:0_max: {}
-      output:
-        111:0_quant: {}
-      attr:
-        output_dtype: u8
-    Add_34:
-      type: InnerProduct
-      input:
-        '576:0': {}
-        111:0_quant: {}
-        bert.encoder.layer.0.attention.self.key.bias:0: {}
-        576:0_min: {}
-        576:0_max: {}
-        111:0_quant_min: {}
-        111:0_quant_max: {}
-        Add_34:0_min: {}
-        Add_34:0_max: {}
-      output:
-        Add_34:0: {}
-      attr:
-        output_dtype: s8
-    Reshape_44:
-      type: Reshape
-      input:
-        Add_34:0: {}
-        input_ids:0: {}
-      output:
-        131:0_quant: {}
-      attr:
-        dst_shape: 4,64,-1,-1
-        dims: '0'
-    Add_46:
-      type: InnerProduct
-      input:
-        '579:0': {}
-        111:0_quant: {}
-        bert.encoder.layer.0.attention.self.value.bias:0: {}
-        579:0_min: {}
-        579:0_max: {}
-        111:0_quant_min: {}
-        111:0_quant_max: {}
-        Add_46:0_min: {}
-        Add_46:0_max: {}
-      output:
-        Add_46:0: {}
-      attr:
-        output_dtype: s8
-    Reshape_56:
-      type: Reshape
-      input:
-        Add_46:0: {}
-        input_ids:0: {}
-      output:
-        148:0_quant: {}
-      attr:
-        dst_shape: 4,64,-1,-1
-        dims: '0'
-    Add_32:
-      type: InnerProduct
-      input:
-        '575:0': {}
-        111:0_quant: {}
-        bert.encoder.layer.0.attention.self.query.bias:0: {}
-        575:0_min: {}
-        575:0_max: {}
-        111:0_quant_min: {}
-        111:0_quant_max: {}
-        Add_32:0_min: {}
-        Add_32:0_max: {}
-      output:
-        Add_32:0: {}
-      attr:
-        output_dtype: s8
-    Reshape_67:
-      type: Reshape
-      input:
-        Add_32:0: {}
-        input_ids:0: {}
-      output:
-        163:0_quant: {}
-      attr:
-        dst_shape: 4,64,-1,-1
-        dims: '0'
-    Add_73:
-      type: Matmul
-      input:
-        163:0_quant: {}
-        131:0_quant: {}
-        padding_sequence:0: {}
-        163:0_quant_min: {}
-        163:0_quant_max: {}
-        131:0_quant_min: {}
-        131:0_quant_max: {}
-        169:0_min: {}
-        169:0_max: {}
-      output:
-        '169:0': {}
-      attr:
-        src0_perm: 2,0,3,1
-        src1_perm: 2,0,1,3
-        output_scale: 0.125
-        format_any: false
-        append_op: binary_add
-    Softmax_74:
-      type: Softmax
-      input:
-        '169:0': {}
-        170:0_quant_min: {}
-        170:0_quant_max: {}
-      output:
-        170:0_quant: {}
-      attr:
-        output_dtype: u8
-    Transpose_76:
-      type: Matmul
-      input:
-        170:0_quant: {}
-        148:0_quant: {}
-        170:0_quant_min: {}
-        170:0_quant_max: {}
-        148:0_quant_min: {}
-        148:0_quant_max: {}
-        172:0_min: {}
-        172:0_max: {}
-      output:
-        '172:0': {}
-      attr:
-        src1_perm: 2,0,3,1
-        dst_perm: 1,3,0,2
-        output_dtype: u8
-    Reshape_86:
-      type: Reshape
-      input:
-        '172:0': {}
-      output:
-        184:0_quant: {}
-      attr:
-        dst_shape: 256,-1
-    Add_89:
-      type: InnerProduct
-      input:
-        '585:0': {}
-        184:0_quant: {}
-        bert.encoder.layer.0.attention.output.dense.bias:0: {}
-        111:0_reorder: {}
-        585:0_min: {}
-        585:0_max: {}
-        184:0_quant_min: {}
-        184:0_quant_max: {}
-        188:0_min: {}
-        188:0_max: {}
-      output:
-        '188:0': {}
-      attr:
-        append_op: sum
-    Add_100:
-      type: LayerNorm
-      input:
-        '188:0': {}
-        bert.encoder.layer.0.attention.output.LayerNorm.weight:0: {}
-        bert.encoder.layer.0.attention.output.LayerNorm.bias:0: {}
-      output:
-        '199:0': {}
-      attr:
-        epsilon: 9.999999960041972e-13
-        transpose_mode: 1,0
-    Mul_110_quant_0:
-      type: Quantize
-      input:
-        '199:0': {}
-        199:0_min: {}
-        199:0_max: {}
-      output:
-        199:0_quant: {}
-      attr:
-        output_dtype: u8
-    Mul_110:
-      type: InnerProduct
-      input:
-        '586:0': {}
-        199:0_quant: {}
-        bert.encoder.layer.0.intermediate.dense.bias:0: {}
-        586:0_min: {}
-        586:0_max: {}
-        199:0_quant_min: {}
-        199:0_quant_max: {}
-        210:0_quant_min: {}
-        210:0_quant_max: {}
-      output:
-        210:0_quant: {}
-    Mul_110_gelu:
-      type: Gelu
-      input:
-        210:0_quant: {}
-      output:
-        210:0_quant_gelu: {}
-      attr:
-        algorithm: gelu_tanh
-    Mul_110_gelu_quant:
-      type: Quantize
-      input:
-        210:0_quant_gelu: {}
-        210:0_quant_min: {}
-        210:0_quant_max: {}
-      output:
-        210:0_quant_quant: {}
-      attr:
-        output_dtype: u8
-    Add_113:
-      type: InnerProduct
-      input:
-        '587:0': {}
-        210:0_quant_quant: {}
-        bert.encoder.layer.0.output.dense.bias:0: {}
-        '199:0': {}
-        587:0_min: {}
-        587:0_max: {}
-        210:0_quant_min: {}
-        210:0_quant_max: {}
-        214:0_min: {}
-        214:0_max: {}
-      output:
-        '214:0': {}
-      attr:
-        append_op: sum
-    Add_124:
-      type: LayerNorm
-      input:
-        '214:0': {}
-        bert.encoder.layer.0.output.LayerNorm.weight:0: {}
-        bert.encoder.layer.0.output.LayerNorm.bias:0: {}
-      output:
-        '225:0': {}
-      attr:
-        epsilon: 9.999999960041972e-13
-        transpose_mode: 1,0
-    Add_128_quant_0:
-      type: Quantize
-      input:
-        '225:0': {}
-        225:0_min: {}
-        225:0_max: {}
-      output:
-        225:0_quant: {}
-      attr:
-        output_dtype: u8
-    Add_128:
-      type: InnerProduct
-      input:
-        '589:0': {}
-        225:0_quant: {}
-        bert.encoder.layer.1.attention.self.key.bias:0: {}
-        589:0_min: {}
-        589:0_max: {}
-        225:0_quant_min: {}
-        225:0_quant_max: {}
-        Add_128:0_min: {}
-        Add_128:0_max: {}
-      output:
-        Add_128:0: {}
-      attr:
-        output_dtype: s8
-    Reshape_138:
-      type: Reshape
-      input:
-        Add_128:0: {}
-        input_ids:0: {}
-      output:
-        245:0_quant: {}
-      attr:
-        dst_shape: 4,64,-1,-1
-        dims: '0'
-    Add_140:
-      type: InnerProduct
-      input:
-        '592:0': {}
-        225:0_quant: {}
-        bert.encoder.layer.1.attention.self.value.bias:0: {}
-        592:0_min: {}
-        592:0_max: {}
-        225:0_quant_min: {}
-        225:0_quant_max: {}
-        Add_140:0_min: {}
-        Add_140:0_max: {}
-      output:
-        Add_140:0: {}
-      attr:
-        output_dtype: s8
-    Reshape_150:
-      type: Reshape
-      input:
-        Add_140:0: {}
-        input_ids:0: {}
-      output:
-        262:0_quant: {}
-      attr:
-        dst_shape: 4,64,-1,-1
-        dims: '0'
-    Add_126:
-      type: InnerProduct
-      input:
-        '588:0': {}
-        225:0_quant: {}
-        bert.encoder.layer.1.attention.self.query.bias:0: {}
-        588:0_min: {}
-        588:0_max: {}
-        225:0_quant_min: {}
-        225:0_quant_max: {}
-        Add_126:0_min: {}
-        Add_126:0_max: {}
-      output:
-        Add_126:0: {}
-      attr:
-        output_dtype: s8
-    Reshape_161:
-      type: Reshape
-      input:
-        Add_126:0: {}
-        input_ids:0: {}
-      output:
-        277:0_quant: {}
-      attr:
-        dst_shape: 4,64,-1,-1
-        dims: '0'
-    Add_167:
-      type: Matmul
-      input:
-        277:0_quant: {}
-        245:0_quant: {}
-        padding_sequence:0: {}
-        277:0_quant_min: {}
-        277:0_quant_max: {}
-        245:0_quant_min: {}
-        245:0_quant_max: {}
-        283:0_min: {}
-        283:0_max: {}
-      output:
-        '283:0': {}
-      attr:
-        src0_perm: 2,0,3,1
-        src1_perm: 2,0,1,3
-        output_scale: 0.125
-        format_any: false
-        append_op: binary_add
-    Softmax_168:
-      type: Softmax
-      input:
-        '283:0': {}
-        284:0_quant_min: {}
-        284:0_quant_max: {}
-      output:
-        284:0_quant: {}
-      attr:
-        output_dtype: u8
-    Transpose_170:
-      type: Matmul
-      input:
-        284:0_quant: {}
-        262:0_quant: {}
-        284:0_quant_min: {}
-        284:0_quant_max: {}
-        262:0_quant_min: {}
-        262:0_quant_max: {}
-        286:0_min: {}
-        286:0_max: {}
-      output:
-        '286:0': {}
-      attr:
-        src1_perm: 2,0,3,1
-        dst_perm: 1,3,0,2
-        output_dtype: u8
-    Reshape_180:
-      type: Reshape
-      input:
-        '286:0': {}
-      output:
-        298:0_quant: {}
-      attr:
-        dst_shape: 256,-1
-    Add_183:
-      type: InnerProduct
-      input:
-        '598:0': {}
-        298:0_quant: {}
-        bert.encoder.layer.1.attention.output.dense.bias:0: {}
-        '225:0': {}
-        598:0_min: {}
-        598:0_max: {}
-        298:0_quant_min: {}
-        298:0_quant_max: {}
-        302:0_min: {}
-        302:0_max: {}
-      output:
-        '302:0': {}
-      attr:
-        append_op: sum
-    Add_194:
-      type: LayerNorm
-      input:
-        '302:0': {}
-        bert.encoder.layer.1.attention.output.LayerNorm.weight:0: {}
-        bert.encoder.layer.1.attention.output.LayerNorm.bias:0: {}
-      output:
-        '313:0': {}
-      attr:
-        epsilon: 9.999999960041972e-13
-        transpose_mode: 1,0
-    Mul_204_quant_0:
-      type: Quantize
-      input:
-        '313:0': {}
-        313:0_min: {}
-        313:0_max: {}
-      output:
-        313:0_quant: {}
-      attr:
-        output_dtype: u8
-    Mul_204:
-      type: InnerProduct
-      input:
-        '599:0': {}
-        313:0_quant: {}
-        bert.encoder.layer.1.intermediate.dense.bias:0: {}
-        599:0_min: {}
-        599:0_max: {}
-        313:0_quant_min: {}
-        313:0_quant_max: {}
-        324:0_quant_min: {}
-        324:0_quant_max: {}
-      output:
-        324:0_quant: {}
-    Mul_204_gelu:
-      type: Gelu
-      input:
-        324:0_quant: {}
-      output:
-        324:0_quant_gelu: {}
-      attr:
-        algorithm: gelu_tanh
-    Mul_204_gelu_quant:
-      type: Quantize
-      input:
-        324:0_quant_gelu: {}
-        324:0_quant_min: {}
-        324:0_quant_max: {}
-      output:
-        324:0_quant_quant: {}
-      attr:
-        output_dtype: u8
-    Add_207:
-      type: InnerProduct
-      input:
-        '600:0': {}
-        324:0_quant_quant: {}
-        bert.encoder.layer.1.output.dense.bias:0: {}
-        '313:0': {}
-        600:0_min: {}
-        600:0_max: {}
-        324:0_quant_min: {}
-        324:0_quant_max: {}
-        328:0_min: {}
-        328:0_max: {}
-      output:
-        '328:0': {}
-      attr:
-        append_op: sum
-    Add_218:
-      type: LayerNorm
-      input:
-        '328:0': {}
-        bert.encoder.layer.1.output.LayerNorm.weight:0: {}
-        bert.encoder.layer.1.output.LayerNorm.bias:0: {}
-      output:
-        '339:0': {}
-      attr:
-        epsilon: 9.999999960041972e-13
-        transpose_mode: 1,0
-    Add_222_quant_0:
-      type: Quantize
-      input:
-        '339:0': {}
-        339:0_min: {}
-        339:0_max: {}
-      output:
-        339:0_quant: {}
-      attr:
-        output_dtype: u8
-    Add_222:
-      type: InnerProduct
-      input:
-        '602:0': {}
-        339:0_quant: {}
-        bert.encoder.layer.2.attention.self.key.bias:0: {}
-        602:0_min: {}
-        602:0_max: {}
-        339:0_quant_min: {}
-        339:0_quant_max: {}
-        Add_222:0_min: {}
-        Add_222:0_max: {}
-      output:
-        Add_222:0: {}
-      attr:
-        output_dtype: s8
-    Reshape_232:
-      type: Reshape
-      input:
-        Add_222:0: {}
-        input_ids:0: {}
-      output:
-        359:0_quant: {}
-      attr:
-        dst_shape: 4,64,-1,-1
-        dims: '0'
-    Add_234:
-      type: InnerProduct
-      input:
-        '605:0': {}
-        339:0_quant: {}
-        bert.encoder.layer.2.attention.self.value.bias:0: {}
-        605:0_min: {}
-        605:0_max: {}
-        339:0_quant_min: {}
-        339:0_quant_max: {}
-        Add_234:0_min: {}
-        Add_234:0_max: {}
-      output:
-        Add_234:0: {}
-      attr:
-        output_dtype: s8
-    Reshape_244:
-      type: Reshape
-      input:
-        Add_234:0: {}
-        input_ids:0: {}
-      output:
-        376:0_quant: {}
-      attr:
-        dst_shape: 4,64,-1,-1
-        dims: '0'
-    Add_220:
-      type: InnerProduct
-      input:
-        '601:0': {}
-        339:0_quant: {}
-        bert.encoder.layer.2.attention.self.query.bias:0: {}
-        601:0_min: {}
-        601:0_max: {}
-        339:0_quant_min: {}
-        339:0_quant_max: {}
-        Add_220:0_min: {}
-        Add_220:0_max: {}
-      output:
-        Add_220:0: {}
-      attr:
-        output_dtype: s8
-    Reshape_255:
-      type: Reshape
-      input:
-        Add_220:0: {}
-        input_ids:0: {}
-      output:
-        391:0_quant: {}
-      attr:
-        dst_shape: 4,64,-1,-1
-        dims: '0'
-    Add_261:
-      type: Matmul
-      input:
-        391:0_quant: {}
-        359:0_quant: {}
-        padding_sequence:0: {}
-        391:0_quant_min: {}
-        391:0_quant_max: {}
-        359:0_quant_min: {}
-        359:0_quant_max: {}
-        397:0_min: {}
-        397:0_max: {}
-      output:
-        '397:0': {}
-      attr:
-        src0_perm: 2,0,3,1
-        src1_perm: 2,0,1,3
-        output_scale: 0.125
-        format_any: false
-        append_op: binary_add
-    Softmax_262:
-      type: Softmax
-      input:
-        '397:0': {}
-        398:0_quant_min: {}
-        398:0_quant_max: {}
-      output:
-        398:0_quant: {}
-      attr:
-        output_dtype: u8
-    Transpose_264:
-      type: Matmul
-      input:
-        398:0_quant: {}
-        376:0_quant: {}
-        398:0_quant_min: {}
-        398:0_quant_max: {}
-        376:0_quant_min: {}
-        376:0_quant_max: {}
-        400:0_min: {}
-        400:0_max: {}
-      output:
-        '400:0': {}
-      attr:
-        src1_perm: 2,0,3,1
-        dst_perm: 1,3,0,2
-        output_dtype: u8
-    Reshape_274:
-      type: Reshape
-      input:
-        '400:0': {}
-      output:
-        412:0_quant: {}
-      attr:
-        dst_shape: 256,-1
-    Add_277:
-      type: InnerProduct
-      input:
-        '611:0': {}
-        412:0_quant: {}
-        bert.encoder.layer.2.attention.output.dense.bias:0: {}
-        '339:0': {}
-        611:0_min: {}
-        611:0_max: {}
-        412:0_quant_min: {}
-        412:0_quant_max: {}
-        416:0_min: {}
-        416:0_max: {}
-      output:
-        '416:0': {}
-      attr:
-        append_op: sum
-    Add_288:
-      type: LayerNorm
-      input:
-        '416:0': {}
-        bert.encoder.layer.2.attention.output.LayerNorm.weight:0: {}
-        bert.encoder.layer.2.attention.output.LayerNorm.bias:0: {}
-      output:
-        '427:0': {}
-      attr:
-        epsilon: 9.999999960041972e-13
-        transpose_mode: 1,0
-    Mul_298_quant_0:
-      type: Quantize
-      input:
-        '427:0': {}
-        427:0_min: {}
-        427:0_max: {}
-      output:
-        427:0_quant: {}
-      attr:
-        output_dtype: u8
-    Mul_298:
-      type: InnerProduct
-      input:
-        '612:0': {}
-        427:0_quant: {}
-        bert.encoder.layer.2.intermediate.dense.bias:0: {}
-        612:0_min: {}
-        612:0_max: {}
-        427:0_quant_min: {}
-        427:0_quant_max: {}
-        438:0_quant_min: {}
-        438:0_quant_max: {}
-      output:
-        438:0_quant: {}
-    Mul_298_gelu:
-      type: Gelu
-      input:
-        438:0_quant: {}
-      output:
-        438:0_quant_gelu: {}
-      attr:
-        algorithm: gelu_tanh
-    Mul_298_gelu_quant:
-      type: Quantize
-      input:
-        438:0_quant_gelu: {}
-        438:0_quant_min: {}
-        438:0_quant_max: {}
-      output:
-        438:0_quant_quant: {}
-      attr:
-        output_dtype: u8
-    Add_301:
-      type: InnerProduct
-      input:
-        '613:0': {}
-        438:0_quant_quant: {}
-        bert.encoder.layer.2.output.dense.bias:0: {}
-        '427:0': {}
-        613:0_min: {}
-        613:0_max: {}
-        438:0_quant_min: {}
-        438:0_quant_max: {}
-        442:0_min: {}
-        442:0_max: {}
-      output:
-        '442:0': {}
-      attr:
-        append_op: sum
-    Add_312:
-      type: LayerNorm
-      input:
-        '442:0': {}
-        bert.encoder.layer.2.output.LayerNorm.weight:0: {}
-        bert.encoder.layer.2.output.LayerNorm.bias:0: {}
-      output:
-        '453:0': {}
-      attr:
-        epsilon: 9.999999960041972e-13
-        transpose_mode: 1,0
-    Add_316_quant_0:
-      type: Quantize
-      input:
-        '453:0': {}
-        453:0_min: {}
-        453:0_max: {}
-      output:
-        453:0_quant: {}
-      attr:
-        output_dtype: u8
-    Add_316:
-      type: InnerProduct
-      input:
-        '615:0': {}
-        453:0_quant: {}
-        bert.encoder.layer.3.attention.self.key.bias:0: {}
-        615:0_min: {}
-        615:0_max: {}
-        453:0_quant_min: {}
-        453:0_quant_max: {}
-        Add_316:0_min: {}
-        Add_316:0_max: {}
-      output:
-        Add_316:0: {}
-      attr:
-        output_dtype: s8
-    Reshape_326:
-      type: Reshape
-      input:
-        Add_316:0: {}
-        input_ids:0: {}
-      output:
-        473:0_quant: {}
-      attr:
-        dst_shape: 4,64,-1,-1
-        dims: '0'
-    Add_328:
-      type: InnerProduct
-      input:
-        '618:0': {}
-        453:0_quant: {}
-        bert.encoder.layer.3.attention.self.value.bias:0: {}
-        618:0_min: {}
-        618:0_max: {}
-        453:0_quant_min: {}
-        453:0_quant_max: {}
-        Add_328:0_min: {}
-        Add_328:0_max: {}
-      output:
-        Add_328:0: {}
-      attr:
-        output_dtype: s8
-    Reshape_338:
-      type: Reshape
-      input:
-        Add_328:0: {}
-        input_ids:0: {}
-      output:
-        490:0_quant: {}
-      attr:
-        dst_shape: 4,64,-1,-1
-        dims: '0'
-    Add_314:
-      type: InnerProduct
-      input:
-        '614:0': {}
-        453:0_quant: {}
-        bert.encoder.layer.3.attention.self.query.bias:0: {}
-        614:0_min: {}
-        614:0_max: {}
-        453:0_quant_min: {}
-        453:0_quant_max: {}
-        Add_314:0_min: {}
-        Add_314:0_max: {}
-      output:
-        Add_314:0: {}
-      attr:
-        output_dtype: s8
-    Reshape_349:
-      type: Reshape
-      input:
-        Add_314:0: {}
-        input_ids:0: {}
-      output:
-        505:0_quant: {}
-      attr:
-        dst_shape: 4,64,-1,-1
-        dims: '0'
-    Add_355:
-      type: Matmul
-      input:
-        505:0_quant: {}
-        473:0_quant: {}
-        padding_sequence:0: {}
-        505:0_quant_min: {}
-        505:0_quant_max: {}
-        473:0_quant_min: {}
-        473:0_quant_max: {}
-        511:0_min: {}
-        511:0_max: {}
-      output:
-        '511:0': {}
-      attr:
-        src0_perm: 2,0,3,1
-        src1_perm: 2,0,1,3
-        output_scale: 0.125
-        format_any: false
-        append_op: binary_add
-    Softmax_356:
-      type: Softmax
-      input:
-        '511:0': {}
-        512:0_quant_min: {}
-        512:0_quant_max: {}
-      output:
-        512:0_quant: {}
-      attr:
-        output_dtype: u8
-    Transpose_358:
-      type: Matmul
-      input:
-        512:0_quant: {}
-        490:0_quant: {}
-        512:0_quant_min: {}
-        512:0_quant_max: {}
-        490:0_quant_min: {}
-        490:0_quant_max: {}
-        514:0_min: {}
-        514:0_max: {}
-      output:
-        '514:0': {}
-      attr:
-        src1_perm: 2,0,3,1
-        dst_perm: 1,3,0,2
-        output_dtype: u8
-    Reshape_368:
-      type: Reshape
-      input:
-        '514:0': {}
-      output:
-        526:0_quant: {}
-      attr:
-        dst_shape: 256,-1
-    Add_371:
-      type: InnerProduct
-      input:
-        '624:0': {}
-        526:0_quant: {}
-        bert.encoder.layer.3.attention.output.dense.bias:0: {}
-        '453:0': {}
-        624:0_min: {}
-        624:0_max: {}
-        526:0_quant_min: {}
-        526:0_quant_max: {}
-        530:0_min: {}
-        530:0_max: {}
-      output:
-        '530:0': {}
-      attr:
-        append_op: sum
-    Add_382:
-      type: LayerNorm
-      input:
-        '530:0': {}
-        bert.encoder.layer.3.attention.output.LayerNorm.weight:0: {}
-        bert.encoder.layer.3.attention.output.LayerNorm.bias:0: {}
-      output:
-        '541:0': {}
-      attr:
-        epsilon: 9.999999960041972e-13
-        transpose_mode: 1,0
-    Mul_392_quant_0:
-      type: Quantize
-      input:
-        '541:0': {}
-        541:0_min: {}
-        541:0_max: {}
-      output:
-        541:0_quant: {}
-      attr:
-        output_dtype: u8
-    Mul_392:
-      type: InnerProduct
-      input:
-        '625:0': {}
-        541:0_quant: {}
-        bert.encoder.layer.3.intermediate.dense.bias:0: {}
-        625:0_min: {}
-        625:0_max: {}
-        541:0_quant_min: {}
-        541:0_quant_max: {}
-        552:0_quant_min: {}
-        552:0_quant_max: {}
-      output:
-        552:0_quant: {}
-    Mul_392_gelu:
-      type: Gelu
-      input:
-        552:0_quant: {}
-      output:
-        552:0_quant_gelu: {}
-      attr:
-        algorithm: gelu_tanh
-    Mul_392_gelu_quant:
-      type: Quantize
-      input:
-        552:0_quant_gelu: {}
-        552:0_quant_min: {}
-        552:0_quant_max: {}
-      output:
-        552:0_quant_quant: {}
-      attr:
-        output_dtype: u8
-    Add_395:
-      type: InnerProduct
-      input:
-        '626:0': {}
-        552:0_quant_quant: {}
-        bert.encoder.layer.3.output.dense.bias:0: {}
-        '541:0': {}
-        626:0_min: {}
-        626:0_max: {}
-        552:0_quant_min: {}
-        552:0_quant_max: {}
-        556:0_min: {}
-        556:0_max: {}
-      output:
-        '556:0': {}
-      attr:
-        append_op: sum
-    Add_406_reorder_pre:
-      type: Reorder
-      input:
-        '556:0': {}
-      output:
-        556:0_reorder: {}
-      attr:
-        src_perm: 0,1
-        dst_perm: 1,0
-    Add_406:
-      type: LayerNorm
-      input:
-        556:0_reorder: {}
-        bert.encoder.layer.3.output.LayerNorm.weight:0: {}
-        bert.encoder.layer.3.output.LayerNorm.bias:0: {}
-      output:
-        Add_406:0: {}
-      attr:
-        epsilon: 9.999999960041972e-13
-    last_layer_reshape:
-      type: Reshape
-      input:
-        Add_406:0: {}
-        input_ids:0: {}
-      output:
-        last_layer_reshape:0: {}
-      attr:
-        dst_shape: -1,-1,256
-        dims: 0,1
-    last_layer_strided_slice:
-      type: StridedSlice
-      input:
-        last_layer_reshape:0: {}
-      output:
-        last_layer_strided_slice:0: {}
-      attr:
-        begin_mask: 5
-        ellipsis_mask: 0
-        end_mask: 5
-        new_axis_mask: 0
-        shrink_axis_mask: 0
-        begin: 0,0,0
-        end: 0,1,0
-        strides: 1,1,1
-    Gather_408:
-      type: Reshape
-      input:
-        last_layer_strided_slice:0: {}
-      output:
-        '569:0': {}
-      attr:
-        dst_shape: -1,256
-    Tanh_410_quant_0:
-      type: Quantize
-      input:
-        '569:0': {}
-        569:0_min: {}
-        569:0_max: {}
-      output:
-        569:0_quant: {}
-      attr:
-        output_dtype: u8
-    Tanh_410:
-      type: InnerProduct
-      input:
-        569:0_quant: {}
-        bert.pooler.dense.weight:0: {}
-        bert.pooler.dense.bias:0: {}
-        569:0_quant_min: {}
-        569:0_quant_max: {}
-        bert.pooler.dense.weight:0_min: {}
-        bert.pooler.dense.weight:0_max: {}
-        571:0_quant_min: {}
-        571:0_quant_max: {}
-      output:
-        571:0_quant: {}
-      attr:
-        src1_perm: 0,1
-        append_op: tanh
-        output_dtype: u8
-    Gemm_411:
-      type: InnerProduct
-      input:
-        571:0_quant: {}
-        classifier.weight:0: {}
-        classifier.bias:0: {}
-        571:0_quant_min: {}
-        571:0_quant_max: {}
-        classifier.weight:0_min: {}
-        classifier.weight:0_max: {}
-        output:0_min: {}
-        output:0_max: {}
-      output:
-        output:0: {}
-      attr:
-        src1_perm: 0,1out
-    output_data:
-      type: Output
-      input:
-        output:0: {}
-        #'199:0': {}
-        #'188:0': {}
-        #184:0_quant: {}

Neural_Engine_INT8_IR/model.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:65046bc15bbc90be9917710793edf85d92dcfd1df7bea5da2c9969fd17327ae4
-size 35125504