add the int8 onnx model and int8 Neural Engine IR

int8-model.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:215393cc42f7417de9fdd919cd9b32ba420f6258f3b9f78f6b2fb4f4948fcf80
+size 35286277

sparse_int8_ir/conf.yaml

@@ -0,0 +1,1866 @@
+model:
+  name: model
+  operator:
+    input_data:
+      type: Input
+      output:
+        input_ids:0:
+          dtype: int32
+          shape: [-1, -1]
+        token_type_ids:0:
+          dtype: int32
+          shape: [-1, -1]
+        attention_mask:0:
+          dtype: int32
+          shape: [-1, -1]
+        bert.embeddings.position_embeddings.weight:0:
+          dtype: fp32
+          shape: [512, 256]
+          location: [0, 524288]
+        bert.embeddings.token_type_embeddings.weight:0:
+          dtype: fp32
+          shape: [2, 256]
+          location: [524288, 2048]
+        bert.embeddings.word_embeddings.weight:0:
+          dtype: fp32
+          shape: [30522, 256]
+          location: [526336, 31254528]
+        bert.embeddings.LayerNorm.weight:0:
+          dtype: fp32
+          shape: [256]
+          location: [31780864, 1024]
+        bert.embeddings.LayerNorm.bias:0:
+          dtype: fp32
+          shape: [256]
+          location: [31781888, 1024]
+        /bert/embeddings/LayerNorm/Add_1_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [31988776, 4]
+        /bert/embeddings/LayerNorm/Add_1_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [31988780, 4]
+        /bert/encoder/layer.0/attention/self/key/Transpose_output_0_quantized: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]
+        /bert/encoder/layer.0/attention/self/key/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [256]
+          location: [31849480, 1024]
+        /bert/encoder/layer.0/attention/self/key/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [256]
+          location: [31850504, 1024]
+        /bert/encoder/layer.0/attention/self/Reshape_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [31988800, 4]
+        /bert/encoder/layer.0/attention/self/Reshape_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [31988804, 4]
+        /bert/encoder/layer.0/attention/self/query/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [256, 256]
+          location: [31851544, 65536]
+        bert.encoder.layer.0.attention.self.query.bias:0:
+          dtype: s32
+          shape: [256]
+          location: [31917080, 1024]
+        /bert/encoder/layer.0/attention/self/query/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [256]
+          location: [31918104, 1024]
+        /bert/encoder/layer.0/attention/self/query/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [256]
+          location: [31919128, 1024]
+        /bert/encoder/layer.0/attention/self/Reshape_2_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [31988792, 4]
+        /bert/encoder/layer.0/attention/self/Reshape_2_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [31988796, 4]
+        /bert/encoder/layer.0/attention/self/value/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [256, 256]
+          location: [31920168, 65536]
+        bert.encoder.layer.0.attention.self.value.bias:0:
+          dtype: s32
+          shape: [256]
+          location: [31985704, 1024]
+        /bert/encoder/layer.0/attention/self/value/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [256]
+          location: [31986728, 1024]
+        /bert/encoder/layer.0/attention/self/value/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [256]
+          location: [31987752, 1024]
+        /bert/encoder/layer.0/attention/self/Reshape_1_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [31988832, 4]
+        /bert/encoder/layer.0/attention/self/Reshape_1_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [31988836, 4]
+        /bert/encoder/layer.0/attention/self/Add_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [31988808, 4]
+        /bert/encoder/layer.0/attention/self/Add_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [31988812, 4]
+        /bert/encoder/layer.0/attention/self/Softmax_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [31988824, 4]
+        /bert/encoder/layer.0/attention/self/Softmax_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [31988828, 4]
+        /bert/encoder/layer.0/attention/self/Reshape_3_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [32057456, 4]
+        /bert/encoder/layer.0/attention/self/Reshape_3_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [32057460, 4]
+        /bert/encoder/layer.0/attention/output/dense/Transpose_output_0_quantized: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]
+        /bert/encoder/layer.0/attention/output/dense/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [256]
+          location: [32055408, 1024]
+        /bert/encoder/layer.0/attention/output/dense/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [256]
+          location: [32056432, 1024]
+        /bert/encoder/layer.0/attention/output/Add_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [32057464, 4]
+        /bert/encoder/layer.0/attention/output/Add_output_0:0_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]
+        /bert/encoder/layer.0/attention/output/LayerNorm/Add_1_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [32333960, 4]
+        /bert/encoder/layer.0/attention/output/LayerNorm/Add_1_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [32333964, 4]
+        /bert/encoder/layer.0/intermediate/dense/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [1024, 256]
+          location: [32059528, 262144]
+        bert.encoder.layer.0.intermediate.dense.bias:0:
+          dtype: s32
+          shape: [1024]
+          location: [32321672, 4096]
+        /bert/encoder/layer.0/intermediate/dense/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [1024]
+          location: [32325768, 4096]
+        /bert/encoder/layer.0/intermediate/dense/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [1024]
+          location: [32329864, 4096]
+        /bert/encoder/layer.0/intermediate/intermediate_act_fn/Mul_1_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [32599192, 4]
+        /bert/encoder/layer.0/intermediate/intermediate_act_fn/Mul_1_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [32599196, 4]
+        /bert/encoder/layer.0/output/dense/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [256, 1024]
+          location: [32333976, 262144]
+        bert.encoder.layer.0.output.dense.bias:0:
+          dtype: s32
+          shape: [256]
+          location: [32596120, 1024]
+        /bert/encoder/layer.0/output/dense/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [256]
+          location: [32597144, 1024]
+        /bert/encoder/layer.0/output/dense/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [256]
+          location: [32598168, 1024]
+        /bert/encoder/layer.0/output/Add_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [32599200, 4]
+        /bert/encoder/layer.0/output/Add_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [32599204, 4]
+        bert.encoder.layer.0.output.LayerNorm.weight:0:
+          dtype: fp32
+          shape: [256]
+          location: [32599208, 1024]
+        bert.encoder.layer.0.output.LayerNorm.bias:0:
+          dtype: fp32
+          shape: [256]
+          location: [32600232, 1024]
+        /bert/encoder/layer.0/output/LayerNorm/Add_1_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [32807120, 4]
+        /bert/encoder/layer.0/output/LayerNorm/Add_1_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [32807124, 4]
+        /bert/encoder/layer.1/attention/self/key/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [256, 256]
+          location: [32601264, 65536]
+        bert.encoder.layer.1.attention.self.key.bias:0:
+          dtype: s32
+          shape: [256]
+          location: [32666800, 1024]
+        /bert/encoder/layer.1/attention/self/key/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [256]
+          location: [32667824, 1024]
+        /bert/encoder/layer.1/attention/self/key/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [256]
+          location: [32668848, 1024]
+        /bert/encoder/layer.1/attention/self/Reshape_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [32807144, 4]
+        /bert/encoder/layer.1/attention/self/Reshape_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [32807148, 4]
+        /bert/encoder/layer.1/attention/self/query/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [256, 256]
+          location: [32669888, 65536]
+        bert.encoder.layer.1.attention.self.query.bias:0:
+          dtype: s32
+          shape: [256]
+          location: [32735424, 1024]
+        /bert/encoder/layer.1/attention/self/query/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [256]
+          location: [32736448, 1024]
+        /bert/encoder/layer.1/attention/self/query/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [256]
+          location: [32737472, 1024]
+        /bert/encoder/layer.1/attention/self/Reshape_2_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [32807136, 4]
+        /bert/encoder/layer.1/attention/self/Reshape_2_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [32807140, 4]
+        /bert/encoder/layer.1/attention/self/value/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [256, 256]
+          location: [32738512, 65536]
+        bert.encoder.layer.1.attention.self.value.bias:0:
+          dtype: s32
+          shape: [256]
+          location: [32804048, 1024]
+        /bert/encoder/layer.1/attention/self/value/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [256]
+          location: [32805072, 1024]
+        /bert/encoder/layer.1/attention/self/value/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [256]
+          location: [32806096, 1024]
+        /bert/encoder/layer.1/attention/self/Reshape_1_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [32807176, 4]
+        /bert/encoder/layer.1/attention/self/Reshape_1_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [32807180, 4]
+        /bert/encoder/layer.1/attention/self/Add_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [32807152, 4]
+        /bert/encoder/layer.1/attention/self/Add_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [32807156, 4]
+        /bert/encoder/layer.1/attention/self/Softmax_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [32807168, 4]
+        /bert/encoder/layer.1/attention/self/Softmax_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [32807172, 4]
+        /bert/encoder/layer.1/attention/self/Reshape_3_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [32875800, 4]
+        /bert/encoder/layer.1/attention/self/Reshape_3_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [32875804, 4]
+        /bert/encoder/layer.1/attention/output/dense/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [256, 256]
+          location: [32807192, 65536]
+        bert.encoder.layer.1.attention.output.dense.bias:0:
+          dtype: s32
+          shape: [256]
+          location: [32872728, 1024]
+        /bert/encoder/layer.1/attention/output/dense/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [256]
+          location: [32873752, 1024]
+        /bert/encoder/layer.1/attention/output/dense/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [256]
+          location: [32874776, 1024]
+        /bert/encoder/layer.1/attention/output/Add_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [32875808, 4]
+        /bert/encoder/layer.1/attention/output/Add_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [32875812, 4]
+        bert.encoder.layer.1.attention.output.LayerNorm.weight:0:
+          dtype: fp32
+          shape: [256]
+          location: [32875816, 1024]
+        bert.encoder.layer.1.attention.output.LayerNorm.bias:0:
+          dtype: fp32
+          shape: [256]
+          location: [32876840, 1024]
+        /bert/encoder/layer.1/attention/output/LayerNorm/Add_1_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [33152304, 4]
+        /bert/encoder/layer.1/attention/output/LayerNorm/Add_1_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [33152308, 4]
+        /bert/encoder/layer.1/intermediate/dense/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [1024, 256]
+          location: [32877872, 262144]
+        bert.encoder.layer.1.intermediate.dense.bias:0:
+          dtype: s32
+          shape: [1024]
+          location: [33140016, 4096]
+        /bert/encoder/layer.1/intermediate/dense/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [1024]
+          location: [33144112, 4096]
+        /bert/encoder/layer.1/intermediate/dense/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [1024]
+          location: [33148208, 4096]
+        /bert/encoder/layer.1/intermediate/intermediate_act_fn/Mul_1_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [33417536, 4]
+        /bert/encoder/layer.1/intermediate/intermediate_act_fn/Mul_1_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [33417540, 4]
+        /bert/encoder/layer.1/output/dense/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [256, 1024]
+          location: [33152320, 262144]
+        bert.encoder.layer.1.output.dense.bias:0:
+          dtype: s32
+          shape: [256]
+          location: [33414464, 1024]
+        /bert/encoder/layer.1/output/dense/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [256]
+          location: [33415488, 1024]
+        /bert/encoder/layer.1/output/dense/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [256]
+          location: [33416512, 1024]
+        /bert/encoder/layer.1/output/Add_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [33417544, 4]
+        /bert/encoder/layer.1/output/Add_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [33417548, 4]
+        bert.encoder.layer.1.output.LayerNorm.weight:0:
+          dtype: fp32
+          shape: [256]
+          location: [33417552, 1024]
+        bert.encoder.layer.1.output.LayerNorm.bias:0:
+          dtype: fp32
+          shape: [256]
+          location: [33418576, 1024]
+        /bert/encoder/layer.1/output/LayerNorm/Add_1_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [33625464, 4]
+        /bert/encoder/layer.1/output/LayerNorm/Add_1_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [33625468, 4]
+        /bert/encoder/layer.2/attention/self/key/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [256, 256]
+          location: [33419608, 65536]
+        bert.encoder.layer.2.attention.self.key.bias:0:
+          dtype: s32
+          shape: [256]
+          location: [33485144, 1024]
+        /bert/encoder/layer.2/attention/self/key/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [256]
+          location: [33486168, 1024]
+        /bert/encoder/layer.2/attention/self/key/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [256]
+          location: [33487192, 1024]
+        /bert/encoder/layer.2/attention/self/Reshape_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [33625488, 4]
+        /bert/encoder/layer.2/attention/self/Reshape_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [33625492, 4]
+        /bert/encoder/layer.2/attention/self/query/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [256, 256]
+          location: [33488232, 65536]
+        bert.encoder.layer.2.attention.self.query.bias:0:
+          dtype: s32
+          shape: [256]
+          location: [33553768, 1024]
+        /bert/encoder/layer.2/attention/self/query/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [256]
+          location: [33554792, 1024]
+        /bert/encoder/layer.2/attention/self/query/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [256]
+          location: [33555816, 1024]
+        /bert/encoder/layer.2/attention/self/Reshape_2_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [33625480, 4]
+        /bert/encoder/layer.2/attention/self/Reshape_2_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [33625484, 4]
+        /bert/encoder/layer.2/attention/self/value/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [256, 256]
+          location: [33556856, 65536]
+        bert.encoder.layer.2.attention.self.value.bias:0:
+          dtype: s32
+          shape: [256]
+          location: [33622392, 1024]
+        /bert/encoder/layer.2/attention/self/value/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [256]
+          location: [33623416, 1024]
+        /bert/encoder/layer.2/attention/self/value/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [256]
+          location: [33624440, 1024]
+        /bert/encoder/layer.2/attention/self/Reshape_1_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [33625520, 4]
+        /bert/encoder/layer.2/attention/self/Reshape_1_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [33625524, 4]
+        /bert/encoder/layer.2/attention/self/Add_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [33625496, 4]
+        /bert/encoder/layer.2/attention/self/Add_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [33625500, 4]
+        /bert/encoder/layer.2/attention/self/Softmax_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [33625512, 4]
+        /bert/encoder/layer.2/attention/self/Softmax_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [33625516, 4]
+        /bert/encoder/layer.2/attention/self/Reshape_3_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [33694144, 4]
+        /bert/encoder/layer.2/attention/self/Reshape_3_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [33694148, 4]
+        /bert/encoder/layer.2/attention/output/dense/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [256, 256]
+          location: [33625536, 65536]
+        bert.encoder.layer.2.attention.output.dense.bias:0:
+          dtype: s32
+          shape: [256]
+          location: [33691072, 1024]
+        /bert/encoder/layer.2/attention/output/dense/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [256]
+          location: [33692096, 1024]
+        /bert/encoder/layer.2/attention/output/dense/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [256]
+          location: [33693120, 1024]
+        /bert/encoder/layer.2/attention/output/Add_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [33694152, 4]
+        /bert/encoder/layer.2/attention/output/Add_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [33694156, 4]
+        bert.encoder.layer.2.attention.output.LayerNorm.weight:0:
+          dtype: fp32
+          shape: [256]
+          location: [33694160, 1024]
+        bert.encoder.layer.2.attention.output.LayerNorm.bias:0:
+          dtype: fp32
+          shape: [256]
+          location: [33695184, 1024]
+        /bert/encoder/layer.2/attention/output/LayerNorm/Add_1_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [33970648, 4]
+        /bert/encoder/layer.2/attention/output/LayerNorm/Add_1_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [33970652, 4]
+        /bert/encoder/layer.2/intermediate/dense/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [1024, 256]
+          location: [33696216, 262144]
+        bert.encoder.layer.2.intermediate.dense.bias:0:
+          dtype: s32
+          shape: [1024]
+          location: [33958360, 4096]
+        /bert/encoder/layer.2/intermediate/dense/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [1024]
+          location: [33962456, 4096]
+        /bert/encoder/layer.2/intermediate/dense/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [1024]
+          location: [33966552, 4096]
+        /bert/encoder/layer.2/intermediate/intermediate_act_fn/Mul_1_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [34235880, 4]
+        /bert/encoder/layer.2/intermediate/intermediate_act_fn/Mul_1_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [34235884, 4]
+        /bert/encoder/layer.2/output/dense/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [256, 1024]
+          location: [33970664, 262144]
+        bert.encoder.layer.2.output.dense.bias:0:
+          dtype: s32
+          shape: [256]
+          location: [34232808, 1024]
+        /bert/encoder/layer.2/output/dense/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [256]
+          location: [34233832, 1024]
+        /bert/encoder/layer.2/output/dense/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [256]
+          location: [34234856, 1024]
+        /bert/encoder/layer.2/output/Add_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [34235888, 4]
+        /bert/encoder/layer.2/output/Add_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [34235892, 4]
+        bert.encoder.layer.2.output.LayerNorm.weight:0:
+          dtype: fp32
+          shape: [256]
+          location: [34235896, 1024]
+        bert.encoder.layer.2.output.LayerNorm.bias:0:
+          dtype: fp32
+          shape: [256]
+          location: [34236920, 1024]
+        /bert/encoder/layer.2/output/LayerNorm/Add_1_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [34443808, 4]
+        /bert/encoder/layer.2/output/LayerNorm/Add_1_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [34443812, 4]
+        /bert/encoder/layer.3/attention/self/key/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [256, 256]
+          location: [34237952, 65536]
+        bert.encoder.layer.3.attention.self.key.bias:0:
+          dtype: s32
+          shape: [256]
+          location: [34303488, 1024]
+        /bert/encoder/layer.3/attention/self/key/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [256]
+          location: [34304512, 1024]
+        /bert/encoder/layer.3/attention/self/key/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [256]
+          location: [34305536, 1024]
+        /bert/encoder/layer.3/attention/self/Reshape_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [34443832, 4]
+        /bert/encoder/layer.3/attention/self/Reshape_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [34443836, 4]
+        /bert/encoder/layer.3/attention/self/query/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [256, 256]
+          location: [34306576, 65536]
+        bert.encoder.layer.3.attention.self.query.bias:0:
+          dtype: s32
+          shape: [256]
+          location: [34372112, 1024]
+        /bert/encoder/layer.3/attention/self/query/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [256]
+          location: [34373136, 1024]
+        /bert/encoder/layer.3/attention/self/query/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [256]
+          location: [34374160, 1024]
+        /bert/encoder/layer.3/attention/self/Reshape_2_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [34443824, 4]
+        /bert/encoder/layer.3/attention/self/Reshape_2_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [34443828, 4]
+        /bert/encoder/layer.3/attention/self/value/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [256, 256]
+          location: [34375200, 65536]
+        bert.encoder.layer.3.attention.self.value.bias:0:
+          dtype: s32
+          shape: [256]
+          location: [34440736, 1024]
+        /bert/encoder/layer.3/attention/self/value/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [256]
+          location: [34441760, 1024]
+        /bert/encoder/layer.3/attention/self/value/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [256]
+          location: [34442784, 1024]
+        /bert/encoder/layer.3/attention/self/Reshape_1_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [34443864, 4]
+        /bert/encoder/layer.3/attention/self/Reshape_1_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [34443868, 4]
+        /bert/encoder/layer.3/attention/self/Add_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [34443840, 4]
+        /bert/encoder/layer.3/attention/self/Add_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [34443844, 4]
+        /bert/encoder/layer.3/attention/self/Softmax_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [34443856, 4]
+        /bert/encoder/layer.3/attention/self/Softmax_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [34443860, 4]
+        /bert/encoder/layer.3/attention/self/Reshape_3_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [34512488, 4]
+        /bert/encoder/layer.3/attention/self/Reshape_3_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [34512492, 4]
+        /bert/encoder/layer.3/attention/output/dense/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [256, 256]
+          location: [34443880, 65536]
+        bert.encoder.layer.3.attention.output.dense.bias:0:
+          dtype: s32
+          shape: [256]
+          location: [34509416, 1024]
+        /bert/encoder/layer.3/attention/output/dense/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [256]
+          location: [34510440, 1024]
+        /bert/encoder/layer.3/attention/output/dense/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [256]
+          location: [34511464, 1024]
+        /bert/encoder/layer.3/attention/output/Add_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [34512496, 4]
+        /bert/encoder/layer.3/attention/output/Add_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [34512500, 4]
+        bert.encoder.layer.3.attention.output.LayerNorm.weight:0:
+          dtype: fp32
+          shape: [256]
+          location: [34512504, 1024]
+        bert.encoder.layer.3.attention.output.LayerNorm.bias:0:
+          dtype: fp32
+          shape: [256]
+          location: [34513528, 1024]
+        /bert/encoder/layer.3/attention/output/LayerNorm/Add_1_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [34788992, 4]
+        /bert/encoder/layer.3/attention/output/LayerNorm/Add_1_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [34788996, 4]
+        /bert/encoder/layer.3/intermediate/dense/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [1024, 256]
+          location: [34514560, 262144]
+        bert.encoder.layer.3.intermediate.dense.bias:0:
+          dtype: s32
+          shape: [1024]
+          location: [34776704, 4096]
+        /bert/encoder/layer.3/intermediate/dense/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [1024]
+          location: [34780800, 4096]
+        /bert/encoder/layer.3/intermediate/dense/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [1024]
+          location: [34784896, 4096]
+        /bert/encoder/layer.3/intermediate/intermediate_act_fn/Mul_1_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [35054224, 4]
+        /bert/encoder/layer.3/intermediate/intermediate_act_fn/Mul_1_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [35054228, 4]
+        /bert/encoder/layer.3/output/dense/Transpose_output_0_quantized:0:
+          dtype: s8
+          shape: [256, 1024]
+          location: [34789008, 262144]
+        bert.encoder.layer.3.output.dense.bias:0:
+          dtype: s32
+          shape: [256]
+          location: [35051152, 1024]
+        /bert/encoder/layer.3/output/dense/Transpose_output_0_quantized:0_min:
+          dtype: fp32
+          shape: [256]
+          location: [35052176, 1024]
+        /bert/encoder/layer.3/output/dense/Transpose_output_0_quantized:0_max:
+          dtype: fp32
+          shape: [256]
+          location: [35053200, 1024]
+        /bert/encoder/layer.3/output/Add_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [35054232, 4]
+        /bert/encoder/layer.3/output/Add_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [35054236, 4]
+        bert.encoder.layer.3.output.LayerNorm.weight:0:
+          dtype: fp32
+          shape: [256]
+          location: [35054240, 1024]
+        bert.encoder.layer.3.output.LayerNorm.bias:0:
+          dtype: fp32
+          shape: [256]
+          location: [35055264, 1024]
+        /bert/pooler/Gather_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [35122856, 4]
+        /bert/pooler/Gather_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [35122860, 4]
+        bert.pooler.dense.weight_quantized:0:
+          dtype: s8
+          shape: [256, 256]
+          location: [35056296, 65536]
+        bert.pooler.dense.bias:0:
+          dtype: s32
+          shape: [256]
+          location: [35121832, 1024]
+        bert.pooler.dense.weight_quantized:0_min:
+          dtype: fp32
+          shape: [256]
+          location: [35122864, 1024]
+        bert.pooler.dense.weight_quantized:0_max:
+          dtype: fp32
+          shape: [256]
+          location: [35123888, 1024]
+        /bert/pooler/activation/Tanh_output_0:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [35125440, 4]
+        /bert/pooler/activation/Tanh_output_0:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [35125444, 4]
+        classifier.weight_quantized:0:
+          dtype: s8
+          shape: [256, 2]
+          location: [35124920, 512]
+        classifier.bias:0:
+          dtype: s32
+          shape: [2]
+          location: [35125432, 8]
+        classifier.weight_quantized:0_min:
+          dtype: fp32
+          shape: [2]
+          location: [35125448, 8]
+        classifier.weight_quantized:0_max:
+          dtype: fp32
+          shape: [2]
+          location: [35125456, 8]
+        609:0_min:
+          dtype: fp32
+          shape: [1]
+          location: [35125464, 4]
+        609:0_max:
+          dtype: fp32
+          shape: [1]
+          location: [35125468, 4]
+    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
+    /bert/embeddings/position_embeddings/Gather:
+      type: Reshape
+      input:
+        position_embeddings/after/reshape:0: {}
+      output:
+        /bert/embeddings/position_embeddings/Gather_output_0:0: {}
+      attr:
+        dst_shape: 1,-1
+    /bert/Mul:
+      type: PaddingSequence
+      input:
+        attention_mask:0: {}
+      output:
+        /bert/Mul_output_0:0: {}
+      attr:
+        dst_shape: -1,4,0,-1
+        dims: 1
+    word_embeddings/reshape:
+      type: Reshape
+      input:
+        input_ids:0: {}
+      output:
+        word_embeddings/reshape:0: {}
+      attr:
+        dst_shape: -1
+    token_type_embeddings/reshape:
+      type: Reshape
+      input:
+        token_type_ids:0: {}
+      output:
+        token_type_embeddings/reshape:0: {}
+      attr:
+        dst_shape: -1
+    /bert/embeddings/token_type_embeddings/Gather:
+      type: Gather
+      input:
+        token_type_embeddings/reshape:0: {}
+        bert.embeddings.token_type_embeddings.weight:0: {}
+        /bert/embeddings/position_embeddings/Gather_output_0:0: {}
+        token_type_ids:0: {}
+      output:
+        /bert/embeddings/token_type_embeddings/Gather:0: {}
+      attr:
+        axis: 0
+        batch_dims: 0
+        append_op: binary_add
+        reshape: -1,-1,256
+        reshape_dims: 0,1
+        mul: 1,2
+    /bert/embeddings/word_embeddings/Gather:
+      type: Gather
+      input:
+        word_embeddings/reshape:0: {}
+        bert.embeddings.word_embeddings.weight:0: {}
+        /bert/embeddings/token_type_embeddings/Gather:0: {}
+        token_type_ids:0: {}
+      output:
+        embeddings_add/reshape_2d:0: {}
+      attr:
+        axis: 0
+        batch_dims: 0
+        append_op: binary_add
+        reshape: -1,-1,256
+        reshape_dims: 0,1
+        mul: 1,2
+    /bert/embeddings/LayerNorm/Add_1:
+      type: LayerNorm
+      input:
+        embeddings_add/reshape_2d:0: {}
+        bert.embeddings.LayerNorm.weight:0: {}
+        bert.embeddings.LayerNorm.bias:0: {}
+      output:
+        /bert/embeddings/LayerNorm/Add_1_output_0:0: {}
+      attr:
+        epsilon: 9.999999960041972e-13
+    /bert/encoder/layer.0/attention/self/key/Add_quant_0_Reorder_Post_0:
+      type: Reorder
+      input:
+        /bert/embeddings/LayerNorm/Add_1_output_0:0: {}
+      output:
+        /bert/embeddings/LayerNorm/Add_1_output_0:0_reorder: {}
+      attr:
+        src_perm: 0,1
+        dst_perm: 1,0
+    /bert/encoder/layer.0/attention/self/key/Add_quant_0:
+      type: Quantize
+      input:
+        /bert/embeddings/LayerNorm/Add_1_output_0:0_reorder: {}
+        /bert/embeddings/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/embeddings/LayerNorm/Add_1_output_0:0_max: {}
+      output:
+        /bert/embeddings/LayerNorm/Add_1_output_0:0_quant: {}
+      attr:
+        output_dtype: u8
+    /bert/encoder/layer.0/attention/self/key/Add:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.0/attention/self/key/Transpose_output_0_quantized:0: {}
+        /bert/embeddings/LayerNorm/Add_1_output_0:0_quant: {}
+        bert.encoder.layer.0.attention.self.key.bias:0: {}
+        /bert/encoder/layer.0/attention/self/key/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.0/attention/self/key/Transpose_output_0_quantized:0_max: {}
+        /bert/embeddings/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/embeddings/LayerNorm/Add_1_output_0:0_max: {}
+        /bert/encoder/layer.0/attention/self/Reshape_output_0:0_min: {}
+        /bert/encoder/layer.0/attention/self/Reshape_output_0:0_max: {}
+        input_ids:0: {}
+      output:
+        /bert/encoder/layer.0/attention/self/Reshape_output_0:0: {}
+      attr:
+        output_dtype: s8
+        reshape: 4,64,-1, -1
+        reshape_dims: '0'
+    /bert/encoder/layer.0/attention/self/query/Add:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.0/attention/self/query/Transpose_output_0_quantized:0: {}
+        /bert/embeddings/LayerNorm/Add_1_output_0:0_quant: {}
+        bert.encoder.layer.0.attention.self.query.bias:0: {}
+        /bert/encoder/layer.0/attention/self/query/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.0/attention/self/query/Transpose_output_0_quantized:0_max: {}
+        /bert/embeddings/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/embeddings/LayerNorm/Add_1_output_0:0_max: {}
+        /bert/encoder/layer.0/attention/self/Reshape_2_output_0:0_min: {}
+        /bert/encoder/layer.0/attention/self/Reshape_2_output_0:0_max: {}
+        input_ids:0: {}
+      output:
+        /bert/encoder/layer.0/attention/self/Reshape_2_output_0:0: {}
+      attr:
+        output_dtype: s8
+        reshape: 4,64,-1, -1
+        reshape_dims: '0'
+    /bert/encoder/layer.0/attention/self/value/Add:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.0/attention/self/value/Transpose_output_0_quantized:0: {}
+        /bert/embeddings/LayerNorm/Add_1_output_0:0_quant: {}
+        bert.encoder.layer.0.attention.self.value.bias:0: {}
+        /bert/encoder/layer.0/attention/self/value/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.0/attention/self/value/Transpose_output_0_quantized:0_max: {}
+        /bert/embeddings/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/embeddings/LayerNorm/Add_1_output_0:0_max: {}
+        /bert/encoder/layer.0/attention/self/Reshape_1_output_0:0_min: {}
+        /bert/encoder/layer.0/attention/self/Reshape_1_output_0:0_max: {}
+        input_ids:0: {}
+      output:
+        /bert/encoder/layer.0/attention/self/Reshape_1_output_0:0: {}
+      attr:
+        output_dtype: s8
+        reshape: 4,64,-1, -1
+        reshape_dims: '0'
+    /bert/encoder/layer.0/attention/self/Add:
+      type: Matmul
+      input:
+        /bert/encoder/layer.0/attention/self/Reshape_2_output_0:0: {}
+        /bert/encoder/layer.0/attention/self/Reshape_output_0:0: {}
+        /bert/Mul_output_0:0: {}
+        /bert/encoder/layer.0/attention/self/Reshape_2_output_0:0_min: {}
+        /bert/encoder/layer.0/attention/self/Reshape_2_output_0:0_max: {}
+        /bert/encoder/layer.0/attention/self/Reshape_output_0:0_min: {}
+        /bert/encoder/layer.0/attention/self/Reshape_output_0:0_max: {}
+        /bert/encoder/layer.0/attention/self/Add_output_0:0_min: {}
+        /bert/encoder/layer.0/attention/self/Add_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.0/attention/self/Add_output_0:0: {}
+      attr:
+        src0_perm: 2,0,3,1
+        src1_perm: 2,0,1,3
+        output_scale: 0.125
+        format_any: false
+        append_op: binary_add
+    /bert/encoder/layer.0/attention/self/Softmax:
+      type: Softmax
+      input:
+        /bert/encoder/layer.0/attention/self/Add_output_0:0: {}
+        /bert/encoder/layer.0/attention/self/Softmax_output_0:0_min: {}
+        /bert/encoder/layer.0/attention/self/Softmax_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.0/attention/self/Softmax_output_0:0: {}
+      attr:
+        output_dtype: u8
+    /bert/encoder/layer.0/attention/self/Transpose_3:
+      type: Matmul
+      input:
+        /bert/encoder/layer.0/attention/self/Softmax_output_0:0: {}
+        /bert/encoder/layer.0/attention/self/Reshape_1_output_0:0: {}
+        /bert/encoder/layer.0/attention/self/Softmax_output_0:0_min: {}
+        /bert/encoder/layer.0/attention/self/Softmax_output_0:0_max: {}
+        /bert/encoder/layer.0/attention/self/Reshape_1_output_0:0_min: {}
+        /bert/encoder/layer.0/attention/self/Reshape_1_output_0:0_max: {}
+        /bert/encoder/layer.0/attention/self/Reshape_3_output_0:0_min: {}
+        /bert/encoder/layer.0/attention/self/Reshape_3_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.0/attention/self/Reshape_3_output_0:0: {}
+      attr:
+        src1_perm: 2,0,3,1
+        dst_perm: 1,3,0,2
+        output_dtype: u8
+        reshape: 256,-1
+    /bert/encoder/layer.0/attention/output/Add:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.0/attention/output/dense/Transpose_output_0_quantized:0: {}
+        /bert/encoder/layer.0/attention/self/Reshape_3_output_0:0: {}
+        bert.encoder.layer.0.attention.output.dense.bias:0: {}
+        /bert/embeddings/LayerNorm/Add_1_output_0:0_reorder: {}
+        /bert/encoder/layer.0/attention/output/dense/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.0/attention/output/dense/Transpose_output_0_quantized:0_max: {}
+        /bert/encoder/layer.0/attention/self/Reshape_3_output_0:0_min: {}
+        /bert/encoder/layer.0/attention/self/Reshape_3_output_0:0_max: {}
+        /bert/encoder/layer.0/attention/output/Add_output_0:0_min: {}
+        /bert/encoder/layer.0/attention/output/Add_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.0/attention/output/Add_output_0:0: {}
+      attr:
+        append_op: sum
+    /bert/encoder/layer.0/attention/output/LayerNorm/Add_1:
+      type: LayerNorm
+      input:
+        /bert/encoder/layer.0/attention/output/Add_output_0:0: {}
+        bert.encoder.layer.0.attention.output.LayerNorm.weight:0: {}
+        bert.encoder.layer.0.attention.output.LayerNorm.bias:0: {}
+      output:
+        /bert/encoder/layer.0/attention/output/LayerNorm/Add_1_output_0:0: {}
+      attr:
+        epsilon: 9.999999960041972e-13
+        transpose_mode: 1, 0
+    /bert/encoder/layer.0/intermediate/intermediate_act_fn/Mul_1_quant_0:
+      type: Quantize
+      input:
+        /bert/encoder/layer.0/attention/output/LayerNorm/Add_1_output_0:0: {}
+        /bert/encoder/layer.0/attention/output/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/encoder/layer.0/attention/output/LayerNorm/Add_1_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.0/attention/output/LayerNorm/Add_1_output_0:0_quant: {}
+      attr:
+        output_dtype: u8
+    /bert/encoder/layer.0/intermediate/intermediate_act_fn/Mul_1:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.0/intermediate/dense/Transpose_output_0_quantized:0: {}
+        /bert/encoder/layer.0/attention/output/LayerNorm/Add_1_output_0:0_quant: {}
+        bert.encoder.layer.0.intermediate.dense.bias:0: {}
+        /bert/encoder/layer.0/intermediate/dense/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.0/intermediate/dense/Transpose_output_0_quantized:0_max: {}
+        /bert/encoder/layer.0/attention/output/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/encoder/layer.0/attention/output/LayerNorm/Add_1_output_0:0_max: {}
+        /bert/encoder/layer.0/intermediate/intermediate_act_fn/Mul_1_output_0:0_min: {}
+        /bert/encoder/layer.0/intermediate/intermediate_act_fn/Mul_1_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.0/intermediate/intermediate_act_fn/Mul_1_output_0:0: {}
+      attr:
+        append_op: gelu_tanh
+        output_dtype: u8
+    /bert/encoder/layer.0/output/Add:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.0/output/dense/Transpose_output_0_quantized:0: {}
+        /bert/encoder/layer.0/intermediate/intermediate_act_fn/Mul_1_output_0:0: {}
+        bert.encoder.layer.0.output.dense.bias:0: {}
+        /bert/encoder/layer.0/attention/output/LayerNorm/Add_1_output_0:0: {}
+        /bert/encoder/layer.0/output/dense/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.0/output/dense/Transpose_output_0_quantized:0_max: {}
+        /bert/encoder/layer.0/intermediate/intermediate_act_fn/Mul_1_output_0:0_min: {}
+        /bert/encoder/layer.0/intermediate/intermediate_act_fn/Mul_1_output_0:0_max: {}
+        /bert/encoder/layer.0/output/Add_output_0:0_min: {}
+        /bert/encoder/layer.0/output/Add_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.0/output/Add_output_0:0: {}
+      attr:
+        append_op: sum
+    /bert/encoder/layer.0/output/LayerNorm/Add_1:
+      type: LayerNorm
+      input:
+        /bert/encoder/layer.0/output/Add_output_0:0: {}
+        bert.encoder.layer.0.output.LayerNorm.weight:0: {}
+        bert.encoder.layer.0.output.LayerNorm.bias:0: {}
+      output:
+        /bert/encoder/layer.0/output/LayerNorm/Add_1_output_0:0: {}
+      attr:
+        epsilon: 9.999999960041972e-13
+        transpose_mode: 1, 0
+    /bert/encoder/layer.1/attention/self/key/Add_quant_0:
+      type: Quantize
+      input:
+        /bert/encoder/layer.0/output/LayerNorm/Add_1_output_0:0: {}
+        /bert/encoder/layer.0/output/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/encoder/layer.0/output/LayerNorm/Add_1_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.0/output/LayerNorm/Add_1_output_0:0_quant: {}
+      attr:
+        output_dtype: u8
+    /bert/encoder/layer.1/attention/self/key/Add:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.1/attention/self/key/Transpose_output_0_quantized:0: {}
+        /bert/encoder/layer.0/output/LayerNorm/Add_1_output_0:0_quant: {}
+        bert.encoder.layer.1.attention.self.key.bias:0: {}
+        /bert/encoder/layer.1/attention/self/key/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.1/attention/self/key/Transpose_output_0_quantized:0_max: {}
+        /bert/encoder/layer.0/output/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/encoder/layer.0/output/LayerNorm/Add_1_output_0:0_max: {}
+        /bert/encoder/layer.1/attention/self/Reshape_output_0:0_min: {}
+        /bert/encoder/layer.1/attention/self/Reshape_output_0:0_max: {}
+        input_ids:0: {}
+      output:
+        /bert/encoder/layer.1/attention/self/Reshape_output_0:0: {}
+      attr:
+        output_dtype: s8
+        reshape: 4,64,-1, -1
+        reshape_dims: '0'
+    /bert/encoder/layer.1/attention/self/query/Add:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.1/attention/self/query/Transpose_output_0_quantized:0: {}
+        /bert/encoder/layer.0/output/LayerNorm/Add_1_output_0:0_quant: {}
+        bert.encoder.layer.1.attention.self.query.bias:0: {}
+        /bert/encoder/layer.1/attention/self/query/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.1/attention/self/query/Transpose_output_0_quantized:0_max: {}
+        /bert/encoder/layer.0/output/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/encoder/layer.0/output/LayerNorm/Add_1_output_0:0_max: {}
+        /bert/encoder/layer.1/attention/self/Reshape_2_output_0:0_min: {}
+        /bert/encoder/layer.1/attention/self/Reshape_2_output_0:0_max: {}
+        input_ids:0: {}
+      output:
+        /bert/encoder/layer.1/attention/self/Reshape_2_output_0:0: {}
+      attr:
+        output_dtype: s8
+        reshape: 4,64,-1, -1
+        reshape_dims: '0'
+    /bert/encoder/layer.1/attention/self/value/Add:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.1/attention/self/value/Transpose_output_0_quantized:0: {}
+        /bert/encoder/layer.0/output/LayerNorm/Add_1_output_0:0_quant: {}
+        bert.encoder.layer.1.attention.self.value.bias:0: {}
+        /bert/encoder/layer.1/attention/self/value/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.1/attention/self/value/Transpose_output_0_quantized:0_max: {}
+        /bert/encoder/layer.0/output/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/encoder/layer.0/output/LayerNorm/Add_1_output_0:0_max: {}
+        /bert/encoder/layer.1/attention/self/Reshape_1_output_0:0_min: {}
+        /bert/encoder/layer.1/attention/self/Reshape_1_output_0:0_max: {}
+        input_ids:0: {}
+      output:
+        /bert/encoder/layer.1/attention/self/Reshape_1_output_0:0: {}
+      attr:
+        output_dtype: s8
+        reshape: 4,64,-1, -1
+        reshape_dims: '0'
+    /bert/encoder/layer.1/attention/self/Add:
+      type: Matmul
+      input:
+        /bert/encoder/layer.1/attention/self/Reshape_2_output_0:0: {}
+        /bert/encoder/layer.1/attention/self/Reshape_output_0:0: {}
+        /bert/Mul_output_0:0: {}
+        /bert/encoder/layer.1/attention/self/Reshape_2_output_0:0_min: {}
+        /bert/encoder/layer.1/attention/self/Reshape_2_output_0:0_max: {}
+        /bert/encoder/layer.1/attention/self/Reshape_output_0:0_min: {}
+        /bert/encoder/layer.1/attention/self/Reshape_output_0:0_max: {}
+        /bert/encoder/layer.1/attention/self/Add_output_0:0_min: {}
+        /bert/encoder/layer.1/attention/self/Add_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.1/attention/self/Add_output_0:0: {}
+      attr:
+        src0_perm: 2,0,3,1
+        src1_perm: 2,0,1,3
+        output_scale: 0.125
+        format_any: false
+        append_op: binary_add
+    /bert/encoder/layer.1/attention/self/Softmax:
+      type: Softmax
+      input:
+        /bert/encoder/layer.1/attention/self/Add_output_0:0: {}
+        /bert/encoder/layer.1/attention/self/Softmax_output_0:0_min: {}
+        /bert/encoder/layer.1/attention/self/Softmax_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.1/attention/self/Softmax_output_0:0: {}
+      attr:
+        output_dtype: u8
+    /bert/encoder/layer.1/attention/self/Transpose_3:
+      type: Matmul
+      input:
+        /bert/encoder/layer.1/attention/self/Softmax_output_0:0: {}
+        /bert/encoder/layer.1/attention/self/Reshape_1_output_0:0: {}
+        /bert/encoder/layer.1/attention/self/Softmax_output_0:0_min: {}
+        /bert/encoder/layer.1/attention/self/Softmax_output_0:0_max: {}
+        /bert/encoder/layer.1/attention/self/Reshape_1_output_0:0_min: {}
+        /bert/encoder/layer.1/attention/self/Reshape_1_output_0:0_max: {}
+        /bert/encoder/layer.1/attention/self/Reshape_3_output_0:0_min: {}
+        /bert/encoder/layer.1/attention/self/Reshape_3_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.1/attention/self/Reshape_3_output_0:0: {}
+      attr:
+        src1_perm: 2,0,3,1
+        dst_perm: 1,3,0,2
+        output_dtype: u8
+        reshape: 256,-1
+    /bert/encoder/layer.1/attention/output/Add:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.1/attention/output/dense/Transpose_output_0_quantized:0: {}
+        /bert/encoder/layer.1/attention/self/Reshape_3_output_0:0: {}
+        bert.encoder.layer.1.attention.output.dense.bias:0: {}
+        /bert/encoder/layer.0/output/LayerNorm/Add_1_output_0:0: {}
+        /bert/encoder/layer.1/attention/output/dense/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.1/attention/output/dense/Transpose_output_0_quantized:0_max: {}
+        /bert/encoder/layer.1/attention/self/Reshape_3_output_0:0_min: {}
+        /bert/encoder/layer.1/attention/self/Reshape_3_output_0:0_max: {}
+        /bert/encoder/layer.1/attention/output/Add_output_0:0_min: {}
+        /bert/encoder/layer.1/attention/output/Add_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.1/attention/output/Add_output_0:0: {}
+      attr:
+        append_op: sum
+    /bert/encoder/layer.1/attention/output/LayerNorm/Add_1:
+      type: LayerNorm
+      input:
+        /bert/encoder/layer.1/attention/output/Add_output_0:0: {}
+        bert.encoder.layer.1.attention.output.LayerNorm.weight:0: {}
+        bert.encoder.layer.1.attention.output.LayerNorm.bias:0: {}
+      output:
+        /bert/encoder/layer.1/attention/output/LayerNorm/Add_1_output_0:0: {}
+      attr:
+        epsilon: 9.999999960041972e-13
+        transpose_mode: 1, 0
+    /bert/encoder/layer.1/intermediate/intermediate_act_fn/Mul_1_quant_0:
+      type: Quantize
+      input:
+        /bert/encoder/layer.1/attention/output/LayerNorm/Add_1_output_0:0: {}
+        /bert/encoder/layer.1/attention/output/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/encoder/layer.1/attention/output/LayerNorm/Add_1_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.1/attention/output/LayerNorm/Add_1_output_0:0_quant: {}
+      attr:
+        output_dtype: u8
+    /bert/encoder/layer.1/intermediate/intermediate_act_fn/Mul_1:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.1/intermediate/dense/Transpose_output_0_quantized:0: {}
+        /bert/encoder/layer.1/attention/output/LayerNorm/Add_1_output_0:0_quant: {}
+        bert.encoder.layer.1.intermediate.dense.bias:0: {}
+        /bert/encoder/layer.1/intermediate/dense/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.1/intermediate/dense/Transpose_output_0_quantized:0_max: {}
+        /bert/encoder/layer.1/attention/output/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/encoder/layer.1/attention/output/LayerNorm/Add_1_output_0:0_max: {}
+        /bert/encoder/layer.1/intermediate/intermediate_act_fn/Mul_1_output_0:0_min: {}
+        /bert/encoder/layer.1/intermediate/intermediate_act_fn/Mul_1_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.1/intermediate/intermediate_act_fn/Mul_1_output_0:0: {}
+      attr:
+        append_op: gelu_tanh
+        output_dtype: u8
+    /bert/encoder/layer.1/output/Add:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.1/output/dense/Transpose_output_0_quantized:0: {}
+        /bert/encoder/layer.1/intermediate/intermediate_act_fn/Mul_1_output_0:0: {}
+        bert.encoder.layer.1.output.dense.bias:0: {}
+        /bert/encoder/layer.1/attention/output/LayerNorm/Add_1_output_0:0: {}
+        /bert/encoder/layer.1/output/dense/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.1/output/dense/Transpose_output_0_quantized:0_max: {}
+        /bert/encoder/layer.1/intermediate/intermediate_act_fn/Mul_1_output_0:0_min: {}
+        /bert/encoder/layer.1/intermediate/intermediate_act_fn/Mul_1_output_0:0_max: {}
+        /bert/encoder/layer.1/output/Add_output_0:0_min: {}
+        /bert/encoder/layer.1/output/Add_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.1/output/Add_output_0:0: {}
+      attr:
+        append_op: sum
+    /bert/encoder/layer.1/output/LayerNorm/Add_1:
+      type: LayerNorm
+      input:
+        /bert/encoder/layer.1/output/Add_output_0:0: {}
+        bert.encoder.layer.1.output.LayerNorm.weight:0: {}
+        bert.encoder.layer.1.output.LayerNorm.bias:0: {}
+      output:
+        /bert/encoder/layer.1/output/LayerNorm/Add_1_output_0:0: {}
+      attr:
+        epsilon: 9.999999960041972e-13
+        transpose_mode: 1, 0
+    /bert/encoder/layer.2/attention/self/key/Add_quant_0:
+      type: Quantize
+      input:
+        /bert/encoder/layer.1/output/LayerNorm/Add_1_output_0:0: {}
+        /bert/encoder/layer.1/output/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/encoder/layer.1/output/LayerNorm/Add_1_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.1/output/LayerNorm/Add_1_output_0:0_quant: {}
+      attr:
+        output_dtype: u8
+    /bert/encoder/layer.2/attention/self/key/Add:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.2/attention/self/key/Transpose_output_0_quantized:0: {}
+        /bert/encoder/layer.1/output/LayerNorm/Add_1_output_0:0_quant: {}
+        bert.encoder.layer.2.attention.self.key.bias:0: {}
+        /bert/encoder/layer.2/attention/self/key/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.2/attention/self/key/Transpose_output_0_quantized:0_max: {}
+        /bert/encoder/layer.1/output/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/encoder/layer.1/output/LayerNorm/Add_1_output_0:0_max: {}
+        /bert/encoder/layer.2/attention/self/Reshape_output_0:0_min: {}
+        /bert/encoder/layer.2/attention/self/Reshape_output_0:0_max: {}
+        input_ids:0: {}
+      output:
+        /bert/encoder/layer.2/attention/self/Reshape_output_0:0: {}
+      attr:
+        output_dtype: s8
+        reshape: 4,64,-1, -1
+        reshape_dims: '0'
+    /bert/encoder/layer.2/attention/self/query/Add:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.2/attention/self/query/Transpose_output_0_quantized:0: {}
+        /bert/encoder/layer.1/output/LayerNorm/Add_1_output_0:0_quant: {}
+        bert.encoder.layer.2.attention.self.query.bias:0: {}
+        /bert/encoder/layer.2/attention/self/query/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.2/attention/self/query/Transpose_output_0_quantized:0_max: {}
+        /bert/encoder/layer.1/output/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/encoder/layer.1/output/LayerNorm/Add_1_output_0:0_max: {}
+        /bert/encoder/layer.2/attention/self/Reshape_2_output_0:0_min: {}
+        /bert/encoder/layer.2/attention/self/Reshape_2_output_0:0_max: {}
+        input_ids:0: {}
+      output:
+        /bert/encoder/layer.2/attention/self/Reshape_2_output_0:0: {}
+      attr:
+        output_dtype: s8
+        reshape: 4,64,-1, -1
+        reshape_dims: '0'
+    /bert/encoder/layer.2/attention/self/value/Add:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.2/attention/self/value/Transpose_output_0_quantized:0: {}
+        /bert/encoder/layer.1/output/LayerNorm/Add_1_output_0:0_quant: {}
+        bert.encoder.layer.2.attention.self.value.bias:0: {}
+        /bert/encoder/layer.2/attention/self/value/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.2/attention/self/value/Transpose_output_0_quantized:0_max: {}
+        /bert/encoder/layer.1/output/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/encoder/layer.1/output/LayerNorm/Add_1_output_0:0_max: {}
+        /bert/encoder/layer.2/attention/self/Reshape_1_output_0:0_min: {}
+        /bert/encoder/layer.2/attention/self/Reshape_1_output_0:0_max: {}
+        input_ids:0: {}
+      output:
+        /bert/encoder/layer.2/attention/self/Reshape_1_output_0:0: {}
+      attr:
+        output_dtype: s8
+        reshape: 4,64,-1, -1
+        reshape_dims: '0'
+    /bert/encoder/layer.2/attention/self/Add:
+      type: Matmul
+      input:
+        /bert/encoder/layer.2/attention/self/Reshape_2_output_0:0: {}
+        /bert/encoder/layer.2/attention/self/Reshape_output_0:0: {}
+        /bert/Mul_output_0:0: {}
+        /bert/encoder/layer.2/attention/self/Reshape_2_output_0:0_min: {}
+        /bert/encoder/layer.2/attention/self/Reshape_2_output_0:0_max: {}
+        /bert/encoder/layer.2/attention/self/Reshape_output_0:0_min: {}
+        /bert/encoder/layer.2/attention/self/Reshape_output_0:0_max: {}
+        /bert/encoder/layer.2/attention/self/Add_output_0:0_min: {}
+        /bert/encoder/layer.2/attention/self/Add_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.2/attention/self/Add_output_0:0: {}
+      attr:
+        src0_perm: 2,0,3,1
+        src1_perm: 2,0,1,3
+        output_scale: 0.125
+        format_any: false
+        append_op: binary_add
+    /bert/encoder/layer.2/attention/self/Softmax:
+      type: Softmax
+      input:
+        /bert/encoder/layer.2/attention/self/Add_output_0:0: {}
+        /bert/encoder/layer.2/attention/self/Softmax_output_0:0_min: {}
+        /bert/encoder/layer.2/attention/self/Softmax_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.2/attention/self/Softmax_output_0:0: {}
+      attr:
+        output_dtype: u8
+    /bert/encoder/layer.2/attention/self/Transpose_3:
+      type: Matmul
+      input:
+        /bert/encoder/layer.2/attention/self/Softmax_output_0:0: {}
+        /bert/encoder/layer.2/attention/self/Reshape_1_output_0:0: {}
+        /bert/encoder/layer.2/attention/self/Softmax_output_0:0_min: {}
+        /bert/encoder/layer.2/attention/self/Softmax_output_0:0_max: {}
+        /bert/encoder/layer.2/attention/self/Reshape_1_output_0:0_min: {}
+        /bert/encoder/layer.2/attention/self/Reshape_1_output_0:0_max: {}
+        /bert/encoder/layer.2/attention/self/Reshape_3_output_0:0_min: {}
+        /bert/encoder/layer.2/attention/self/Reshape_3_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.2/attention/self/Reshape_3_output_0:0: {}
+      attr:
+        src1_perm: 2,0,3,1
+        dst_perm: 1,3,0,2
+        output_dtype: u8
+        reshape: 256,-1
+    /bert/encoder/layer.2/attention/output/Add:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.2/attention/output/dense/Transpose_output_0_quantized:0: {}
+        /bert/encoder/layer.2/attention/self/Reshape_3_output_0:0: {}
+        bert.encoder.layer.2.attention.output.dense.bias:0: {}
+        /bert/encoder/layer.1/output/LayerNorm/Add_1_output_0:0: {}
+        /bert/encoder/layer.2/attention/output/dense/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.2/attention/output/dense/Transpose_output_0_quantized:0_max: {}
+        /bert/encoder/layer.2/attention/self/Reshape_3_output_0:0_min: {}
+        /bert/encoder/layer.2/attention/self/Reshape_3_output_0:0_max: {}
+        /bert/encoder/layer.2/attention/output/Add_output_0:0_min: {}
+        /bert/encoder/layer.2/attention/output/Add_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.2/attention/output/Add_output_0:0: {}
+      attr:
+        append_op: sum
+    /bert/encoder/layer.2/attention/output/LayerNorm/Add_1:
+      type: LayerNorm
+      input:
+        /bert/encoder/layer.2/attention/output/Add_output_0:0: {}
+        bert.encoder.layer.2.attention.output.LayerNorm.weight:0: {}
+        bert.encoder.layer.2.attention.output.LayerNorm.bias:0: {}
+      output:
+        /bert/encoder/layer.2/attention/output/LayerNorm/Add_1_output_0:0: {}
+      attr:
+        epsilon: 9.999999960041972e-13
+        transpose_mode: 1, 0
+    /bert/encoder/layer.2/intermediate/intermediate_act_fn/Mul_1_quant_0:
+      type: Quantize
+      input:
+        /bert/encoder/layer.2/attention/output/LayerNorm/Add_1_output_0:0: {}
+        /bert/encoder/layer.2/attention/output/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/encoder/layer.2/attention/output/LayerNorm/Add_1_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.2/attention/output/LayerNorm/Add_1_output_0:0_quant: {}
+      attr:
+        output_dtype: u8
+    /bert/encoder/layer.2/intermediate/intermediate_act_fn/Mul_1:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.2/intermediate/dense/Transpose_output_0_quantized:0: {}
+        /bert/encoder/layer.2/attention/output/LayerNorm/Add_1_output_0:0_quant: {}
+        bert.encoder.layer.2.intermediate.dense.bias:0: {}
+        /bert/encoder/layer.2/intermediate/dense/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.2/intermediate/dense/Transpose_output_0_quantized:0_max: {}
+        /bert/encoder/layer.2/attention/output/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/encoder/layer.2/attention/output/LayerNorm/Add_1_output_0:0_max: {}
+        /bert/encoder/layer.2/intermediate/intermediate_act_fn/Mul_1_output_0:0_min: {}
+        /bert/encoder/layer.2/intermediate/intermediate_act_fn/Mul_1_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.2/intermediate/intermediate_act_fn/Mul_1_output_0:0: {}
+      attr:
+        append_op: gelu_tanh
+        output_dtype: u8
+    /bert/encoder/layer.2/output/Add:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.2/output/dense/Transpose_output_0_quantized:0: {}
+        /bert/encoder/layer.2/intermediate/intermediate_act_fn/Mul_1_output_0:0: {}
+        bert.encoder.layer.2.output.dense.bias:0: {}
+        /bert/encoder/layer.2/attention/output/LayerNorm/Add_1_output_0:0: {}
+        /bert/encoder/layer.2/output/dense/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.2/output/dense/Transpose_output_0_quantized:0_max: {}
+        /bert/encoder/layer.2/intermediate/intermediate_act_fn/Mul_1_output_0:0_min: {}
+        /bert/encoder/layer.2/intermediate/intermediate_act_fn/Mul_1_output_0:0_max: {}
+        /bert/encoder/layer.2/output/Add_output_0:0_min: {}
+        /bert/encoder/layer.2/output/Add_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.2/output/Add_output_0:0: {}
+      attr:
+        append_op: sum
+    /bert/encoder/layer.2/output/LayerNorm/Add_1:
+      type: LayerNorm
+      input:
+        /bert/encoder/layer.2/output/Add_output_0:0: {}
+        bert.encoder.layer.2.output.LayerNorm.weight:0: {}
+        bert.encoder.layer.2.output.LayerNorm.bias:0: {}
+      output:
+        /bert/encoder/layer.2/output/LayerNorm/Add_1_output_0:0: {}
+      attr:
+        epsilon: 9.999999960041972e-13
+        transpose_mode: 1, 0
+    /bert/encoder/layer.3/attention/self/key/Add_quant_0:
+      type: Quantize
+      input:
+        /bert/encoder/layer.2/output/LayerNorm/Add_1_output_0:0: {}
+        /bert/encoder/layer.2/output/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/encoder/layer.2/output/LayerNorm/Add_1_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.2/output/LayerNorm/Add_1_output_0:0_quant: {}
+      attr:
+        output_dtype: u8
+    /bert/encoder/layer.3/attention/self/key/Add:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.3/attention/self/key/Transpose_output_0_quantized:0: {}
+        /bert/encoder/layer.2/output/LayerNorm/Add_1_output_0:0_quant: {}
+        bert.encoder.layer.3.attention.self.key.bias:0: {}
+        /bert/encoder/layer.3/attention/self/key/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.3/attention/self/key/Transpose_output_0_quantized:0_max: {}
+        /bert/encoder/layer.2/output/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/encoder/layer.2/output/LayerNorm/Add_1_output_0:0_max: {}
+        /bert/encoder/layer.3/attention/self/Reshape_output_0:0_min: {}
+        /bert/encoder/layer.3/attention/self/Reshape_output_0:0_max: {}
+        input_ids:0: {}
+      output:
+        /bert/encoder/layer.3/attention/self/Reshape_output_0:0: {}
+      attr:
+        output_dtype: s8
+        reshape: 4,64,-1, -1
+        reshape_dims: '0'
+    /bert/encoder/layer.3/attention/self/query/Add:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.3/attention/self/query/Transpose_output_0_quantized:0: {}
+        /bert/encoder/layer.2/output/LayerNorm/Add_1_output_0:0_quant: {}
+        bert.encoder.layer.3.attention.self.query.bias:0: {}
+        /bert/encoder/layer.3/attention/self/query/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.3/attention/self/query/Transpose_output_0_quantized:0_max: {}
+        /bert/encoder/layer.2/output/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/encoder/layer.2/output/LayerNorm/Add_1_output_0:0_max: {}
+        /bert/encoder/layer.3/attention/self/Reshape_2_output_0:0_min: {}
+        /bert/encoder/layer.3/attention/self/Reshape_2_output_0:0_max: {}
+        input_ids:0: {}
+      output:
+        /bert/encoder/layer.3/attention/self/Reshape_2_output_0:0: {}
+      attr:
+        output_dtype: s8
+        reshape: 4,64,-1, -1
+        reshape_dims: '0'
+    /bert/encoder/layer.3/attention/self/value/Add:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.3/attention/self/value/Transpose_output_0_quantized:0: {}
+        /bert/encoder/layer.2/output/LayerNorm/Add_1_output_0:0_quant: {}
+        bert.encoder.layer.3.attention.self.value.bias:0: {}
+        /bert/encoder/layer.3/attention/self/value/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.3/attention/self/value/Transpose_output_0_quantized:0_max: {}
+        /bert/encoder/layer.2/output/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/encoder/layer.2/output/LayerNorm/Add_1_output_0:0_max: {}
+        /bert/encoder/layer.3/attention/self/Reshape_1_output_0:0_min: {}
+        /bert/encoder/layer.3/attention/self/Reshape_1_output_0:0_max: {}
+        input_ids:0: {}
+      output:
+        /bert/encoder/layer.3/attention/self/Reshape_1_output_0:0: {}
+      attr:
+        output_dtype: s8
+        reshape: 4,64,-1, -1
+        reshape_dims: '0'
+    /bert/encoder/layer.3/attention/self/Add:
+      type: Matmul
+      input:
+        /bert/encoder/layer.3/attention/self/Reshape_2_output_0:0: {}
+        /bert/encoder/layer.3/attention/self/Reshape_output_0:0: {}
+        /bert/Mul_output_0:0: {}
+        /bert/encoder/layer.3/attention/self/Reshape_2_output_0:0_min: {}
+        /bert/encoder/layer.3/attention/self/Reshape_2_output_0:0_max: {}
+        /bert/encoder/layer.3/attention/self/Reshape_output_0:0_min: {}
+        /bert/encoder/layer.3/attention/self/Reshape_output_0:0_max: {}
+        /bert/encoder/layer.3/attention/self/Add_output_0:0_min: {}
+        /bert/encoder/layer.3/attention/self/Add_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.3/attention/self/Add_output_0:0: {}
+      attr:
+        src0_perm: 2,0,3,1
+        src1_perm: 2,0,1,3
+        output_scale: 0.125
+        format_any: false
+        append_op: binary_add
+    /bert/encoder/layer.3/attention/self/Softmax:
+      type: Softmax
+      input:
+        /bert/encoder/layer.3/attention/self/Add_output_0:0: {}
+        /bert/encoder/layer.3/attention/self/Softmax_output_0:0_min: {}
+        /bert/encoder/layer.3/attention/self/Softmax_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.3/attention/self/Softmax_output_0:0: {}
+      attr:
+        output_dtype: u8
+    /bert/encoder/layer.3/attention/self/Transpose_3:
+      type: Matmul
+      input:
+        /bert/encoder/layer.3/attention/self/Softmax_output_0:0: {}
+        /bert/encoder/layer.3/attention/self/Reshape_1_output_0:0: {}
+        /bert/encoder/layer.3/attention/self/Softmax_output_0:0_min: {}
+        /bert/encoder/layer.3/attention/self/Softmax_output_0:0_max: {}
+        /bert/encoder/layer.3/attention/self/Reshape_1_output_0:0_min: {}
+        /bert/encoder/layer.3/attention/self/Reshape_1_output_0:0_max: {}
+        /bert/encoder/layer.3/attention/self/Reshape_3_output_0:0_min: {}
+        /bert/encoder/layer.3/attention/self/Reshape_3_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.3/attention/self/Reshape_3_output_0:0: {}
+      attr:
+        src1_perm: 2,0,3,1
+        dst_perm: 1,3,0,2
+        output_dtype: u8
+        reshape: 256,-1
+    /bert/encoder/layer.3/attention/output/Add:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.3/attention/output/dense/Transpose_output_0_quantized:0: {}
+        /bert/encoder/layer.3/attention/self/Reshape_3_output_0:0: {}
+        bert.encoder.layer.3.attention.output.dense.bias:0: {}
+        /bert/encoder/layer.2/output/LayerNorm/Add_1_output_0:0: {}
+        /bert/encoder/layer.3/attention/output/dense/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.3/attention/output/dense/Transpose_output_0_quantized:0_max: {}
+        /bert/encoder/layer.3/attention/self/Reshape_3_output_0:0_min: {}
+        /bert/encoder/layer.3/attention/self/Reshape_3_output_0:0_max: {}
+        /bert/encoder/layer.3/attention/output/Add_output_0:0_min: {}
+        /bert/encoder/layer.3/attention/output/Add_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.3/attention/output/Add_output_0:0: {}
+      attr:
+        append_op: sum
+    /bert/encoder/layer.3/attention/output/LayerNorm/Add_1:
+      type: LayerNorm
+      input:
+        /bert/encoder/layer.3/attention/output/Add_output_0:0: {}
+        bert.encoder.layer.3.attention.output.LayerNorm.weight:0: {}
+        bert.encoder.layer.3.attention.output.LayerNorm.bias:0: {}
+      output:
+        /bert/encoder/layer.3/attention/output/LayerNorm/Add_1_output_0:0: {}
+      attr:
+        epsilon: 9.999999960041972e-13
+        transpose_mode: 1, 0
+    /bert/encoder/layer.3/intermediate/intermediate_act_fn/Mul_1_quant_0:
+      type: Quantize
+      input:
+        /bert/encoder/layer.3/attention/output/LayerNorm/Add_1_output_0:0: {}
+        /bert/encoder/layer.3/attention/output/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/encoder/layer.3/attention/output/LayerNorm/Add_1_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.3/attention/output/LayerNorm/Add_1_output_0:0_quant: {}
+      attr:
+        output_dtype: u8
+    /bert/encoder/layer.3/intermediate/intermediate_act_fn/Mul_1:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.3/intermediate/dense/Transpose_output_0_quantized:0: {}
+        /bert/encoder/layer.3/attention/output/LayerNorm/Add_1_output_0:0_quant: {}
+        bert.encoder.layer.3.intermediate.dense.bias:0: {}
+        /bert/encoder/layer.3/intermediate/dense/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.3/intermediate/dense/Transpose_output_0_quantized:0_max: {}
+        /bert/encoder/layer.3/attention/output/LayerNorm/Add_1_output_0:0_min: {}
+        /bert/encoder/layer.3/attention/output/LayerNorm/Add_1_output_0:0_max: {}
+        /bert/encoder/layer.3/intermediate/intermediate_act_fn/Mul_1_output_0:0_min: {}
+        /bert/encoder/layer.3/intermediate/intermediate_act_fn/Mul_1_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.3/intermediate/intermediate_act_fn/Mul_1_output_0:0: {}
+      attr:
+        append_op: gelu_tanh
+        output_dtype: u8
+    /bert/encoder/layer.3/output/Add:
+      type: InnerProduct
+      input:
+        /bert/encoder/layer.3/output/dense/Transpose_output_0_quantized:0: {}
+        /bert/encoder/layer.3/intermediate/intermediate_act_fn/Mul_1_output_0:0: {}
+        bert.encoder.layer.3.output.dense.bias:0: {}
+        /bert/encoder/layer.3/attention/output/LayerNorm/Add_1_output_0:0: {}
+        /bert/encoder/layer.3/output/dense/Transpose_output_0_quantized:0_min: {}
+        /bert/encoder/layer.3/output/dense/Transpose_output_0_quantized:0_max: {}
+        /bert/encoder/layer.3/intermediate/intermediate_act_fn/Mul_1_output_0:0_min: {}
+        /bert/encoder/layer.3/intermediate/intermediate_act_fn/Mul_1_output_0:0_max: {}
+        /bert/encoder/layer.3/output/Add_output_0:0_min: {}
+        /bert/encoder/layer.3/output/Add_output_0:0_max: {}
+      output:
+        /bert/encoder/layer.3/output/Add_output_0:0: {}
+      attr:
+        append_op: sum
+    /bert/encoder/layer.3/output/Add_Reorder_Recover:
+      type: Reorder
+      input:
+        /bert/encoder/layer.3/output/Add_output_0:0: {}
+      output:
+        /bert/encoder/layer.3/output/Add_output_0:0_recover: {}
+      attr:
+        src_perm: 0,1
+        dst_perm: 1,0
+    /bert/encoder/layer.3/output/LayerNorm/Add_1:
+      type: LayerNorm
+      input:
+        /bert/encoder/layer.3/output/Add_output_0:0_recover: {}
+        bert.encoder.layer.3.output.LayerNorm.weight:0: {}
+        bert.encoder.layer.3.output.LayerNorm.bias:0: {}
+      output:
+        /bert/encoder/layer.3/output/LayerNorm/Add_1:0: {}
+      attr:
+        epsilon: 9.999999960041972e-13
+    last_layer_reshape:
+      type: Reshape
+      input:
+        /bert/encoder/layer.3/output/LayerNorm/Add_1:0: {}
+        input_ids:0: {}
+      output:
+        last_layer_reshape:0: {}
+      attr:
+        dst_shape: -1,-1,256
+        dims: '0'
+    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
+    /bert/pooler/Gather:
+      type: Reshape
+      input:
+        last_layer_strided_slice:0: {}
+      output:
+        /bert/pooler/Gather_output_0:0: {}
+      attr:
+        dst_shape: -1,256
+    /bert/pooler/activation/Tanh_quant_0:
+      type: Quantize
+      input:
+        /bert/pooler/Gather_output_0:0: {}
+        /bert/pooler/Gather_output_0:0_min: {}
+        /bert/pooler/Gather_output_0:0_max: {}
+      output:
+        /bert/pooler/Gather_output_0:0_quant: {}
+      attr:
+        output_dtype: u8
+    /bert/pooler/activation/Tanh:
+      type: InnerProduct
+      input:
+        /bert/pooler/Gather_output_0:0_quant: {}
+        bert.pooler.dense.weight_quantized:0: {}
+        bert.pooler.dense.bias:0: {}
+        /bert/pooler/Gather_output_0:0_min: {}
+        /bert/pooler/Gather_output_0:0_max: {}
+        bert.pooler.dense.weight_quantized:0_min: {}
+        bert.pooler.dense.weight_quantized:0_max: {}
+        /bert/pooler/activation/Tanh_output_0:0_min: {}
+        /bert/pooler/activation/Tanh_output_0:0_max: {}
+      output:
+        /bert/pooler/activation/Tanh_output_0:0: {}
+      attr:
+        src1_perm: 1,0
+        append_op: tanh
+        output_dtype: u8
+    /classifier/Gemm_Add:
+      type: InnerProduct
+      input:
+        /bert/pooler/activation/Tanh_output_0:0: {}
+        classifier.weight_quantized:0: {}
+        classifier.bias:0: {}
+        /bert/pooler/activation/Tanh_output_0:0_min: {}
+        /bert/pooler/activation/Tanh_output_0:0_max: {}
+        classifier.weight_quantized:0_min: {}
+        classifier.weight_quantized:0_max: {}
+        609:0_min: {}
+        609:0_max: {}
+      output:
+        '609:0': {}
+      attr:
+        src1_perm: 1,0
+    output_data:
+      type: Output
+      input:
+        '609:0': {}

sparse_int8_ir/model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1a7ee7fd4cc774c0662a7842654226c9fc916419cb11c8a8bf46673267f9b13
+size 35125472