<
   ir_version: 10,
   opset_import: ["" : 23],
   producer_name: "pytorch"
>
main_graph (int32[batch,77] text) => (float[batch,512] text_embedding) 
   <
      float[batch,77,768] add_1050
      float[batch,77,768] add_1075
      float[batch,77,768] add_1148
      float[batch,77,768] add_1173
      float[batch,77,768] add_1246
      float[batch,77,768] add_1271
      float[batch,77,768] add_168
      float[batch,77,768] add_193
      float[batch,77,768] add_266
      float[batch,77,768] add_291
      float[batch,77,768] add_364
      float[batch,77,768] add_389
      float[batch,77,768] add_46
      float[batch,77,768] add_462
      float[batch,77,768] add_487
      float[batch,77,768] add_55
      float[batch,77,768] add_560
      float[batch,77,768] add_585
      float[batch,77,768] add_658
      float[batch,77,768] add_683
      float[batch,77,768] add_756
      float[batch,77,768] add_781
      float[batch,77,768] add_854
      float[batch,77,768] add_879
      float[batch,77,768] add_952
      float[batch,77,768] add_977
      float[batch,1,1,77] bitwise_and_1_f
      float[batch,1] clamp_min
      float[batch,768] div
      float[batch,77,768] embedding
      float[batch,77,3072] gelu
      float[batch,77,3072] gelu_1
      float[batch,77,3072] gelu_10
      float[batch,77,3072] gelu_11
      float[batch,640] gelu_12
      float[batch,77,3072] gelu_2
      float[batch,77,3072] gelu_3
      float[batch,77,3072] gelu_4
      float[batch,77,3072] gelu_5
      float[batch,77,3072] gelu_6
      float[batch,77,3072] gelu_7
      float[batch,77,3072] gelu_8
      float[batch,77,3072] gelu_9
      float[batch,77,768] layer_norm
      float[batch,77,768] layer_norm_1
      float[batch,77,768] layer_norm_10
      float[batch,77,768] layer_norm_11
      float[batch,77,768] layer_norm_12
      float[batch,77,768] layer_norm_13
      float[batch,77,768] layer_norm_14
      float[batch,77,768] layer_norm_15
      float[batch,77,768] layer_norm_16
      float[batch,77,768] layer_norm_17
      float[batch,77,768] layer_norm_18
      float[batch,77,768] layer_norm_19
      float[batch,77,768] layer_norm_2
      float[batch,77,768] layer_norm_20
      float[batch,77,768] layer_norm_21
      float[batch,77,768] layer_norm_22
      float[batch,77,768] layer_norm_23
      float[batch,77,768] layer_norm_24
      float[batch,77,768] layer_norm_3
      float[batch,77,768] layer_norm_4
      float[batch,77,768] layer_norm_5
      float[batch,77,768] layer_norm_6
      float[batch,77,768] layer_norm_7
      float[batch,77,768] layer_norm_8
      float[batch,77,768] layer_norm_9
      float[batch,1] linalg_vector_norm
      float[batch,77,768] linear
      float[batch,77,768] linear_1
      float[batch,77,3072] linear_10
      float[batch,77,768] linear_11
      float[batch,77,768] linear_12
      float[batch,77,768] linear_13
      float[batch,77,768] linear_15
      float[batch,77,3072] linear_16
      float[batch,77,768] linear_17
      float[batch,77,768] linear_18
      float[batch,77,768] linear_19
      float[batch,77,768] linear_21
      float[batch,77,3072] linear_22
      float[batch,77,768] linear_23
      float[batch,77,768] linear_24
      float[batch,77,768] linear_25
      float[batch,77,768] linear_27
      float[batch,77,3072] linear_28
      float[batch,77,768] linear_29
      float[batch,77,768] linear_3
      float[batch,77,768] linear_30
      float[batch,77,768] linear_31
      float[batch,77,768] linear_33
      float[batch,77,3072] linear_34
      float[batch,77,768] linear_35
      float[batch,77,768] linear_36
      float[batch,77,768] linear_37
      float[batch,77,768] linear_39
      float[batch,77,3072] linear_4
      float[batch,77,3072] linear_40
      float[batch,77,768] linear_41
      float[batch,77,768] linear_42
      float[batch,77,768] linear_43
      float[batch,77,768] linear_45
      float[batch,77,3072] linear_46
      float[batch,77,768] linear_47
      float[batch,77,768] linear_48
      float[batch,77,768] linear_49
      float[batch,77,768] linear_5
      float[batch,77,768] linear_51
      float[batch,77,3072] linear_52
      float[batch,77,768] linear_53
      float[batch,77,768] linear_54
      float[batch,77,768] linear_55
      float[batch,77,768] linear_57
      float[batch,77,3072] linear_58
      float[batch,77,768] linear_59
      float[batch,77,768] linear_6
      float[batch,77,768] linear_60
      float[batch,77,768] linear_61
      float[batch,77,768] linear_63
      float[batch,77,3072] linear_64
      float[batch,77,768] linear_65
      float[batch,77,768] linear_66
      float[batch,77,768] linear_67
      float[batch,77,768] linear_69
      float[batch,77,768] linear_7
      float[batch,77,3072] linear_70
      float[batch,77,768] linear_71
      float[batch,640] linear_72
      float[batch,512] linear_73
      float[batch,77,768] linear_9
      float[batch,77,768] mul_637
      float[batch,77,768] scaled_dot_product_attention
      float[batch,77,768] scaled_dot_product_attention_1
      float[batch,77,768] scaled_dot_product_attention_10
      float[batch,77,768] scaled_dot_product_attention_11
      float[batch,77,768] scaled_dot_product_attention_2
      float[batch,77,768] scaled_dot_product_attention_3
      float[batch,77,768] scaled_dot_product_attention_4
      float[batch,77,768] scaled_dot_product_attention_5
      float[batch,77,768] scaled_dot_product_attention_6
      float[batch,77,768] scaled_dot_product_attention_7
      float[batch,77,768] scaled_dot_product_attention_8
      float[batch,77,768] scaled_dot_product_attention_9
      float[batch,768] sum_1
      float[batch,1] sum_2_f
      float[batch,77] text_keep
      float[batch,77,1] unsqueeze_12_f
      float[batch,77,768] val_100
      float[batch,77,768] val_101
      float[batch,77,768] val_102
      float[batch,77,768] val_103
      float[batch,77,3072] val_104
      float[batch,77,768] val_105
      float[batch,77,768] val_106
      float[batch,77,768] val_107
      float[batch,77,768] val_108
      float[batch,77,768] val_109
      float[batch,77,3072] val_110
      float[batch,77,768] val_111
      float[batch,77,768] val_112
      float[batch,77,768] val_113
      float[batch,77,768] val_114
      float[batch,77,768] val_115
      float[batch,77,3072] val_116
      float[batch,77,768] val_117
      float[batch,77,768] val_118
      float[batch,77,768] val_119
      float[batch,77,768] val_120
      float[batch,77,768] val_121
      float[batch,77,3072] val_122
      float[batch,77,768] val_123
      float[batch,77,768] val_124
      float[batch,77,768] val_125
      float[batch,77,768] val_126
      float[batch,77,768] val_127
      float[batch,77,3072] val_128
      float[batch,77,768] val_129
      float[batch,77,768] val_130
      float[batch,77,768] val_131
      float[batch,77,768] val_132
      float[batch,77,768] val_133
      float[batch,77,3072] val_134
      float[batch,77,768] val_135
      float[batch,77,768] val_136
      float[batch,77,768] val_137
      float[batch,77,768] val_138
      float[batch,77,768] val_139
      float[batch,77,3072] val_140
      float[batch,77,768] val_141
      float[batch,77,768] val_142
      float[batch,77,768] val_143
      float[batch,77,768] val_144
      float[batch,77,768] val_145
      float[batch,77,3072] val_146
      float[batch,77,768] val_147
      float[batch,1,1,77] val_52_f
      float[batch,1,1,77] val_52_f_bias
      float[batch,1,77,77] val_52_f_mask
      float[batch,77,768] val_76
      float[batch,77,768] val_77
      float[batch,77,768] val_78
      float[batch,77,768] val_79
      float[batch,77,3072] val_80
      float[batch,77,768] val_81
      float[batch,77,768] val_82
      float[batch,77,768] val_83
      float[batch,77,768] val_84
      float[batch,77,768] val_85
      float[batch,77,3072] val_86
      float[batch,77,768] val_87
      float[batch,77,768] val_88
      float[batch,77,768] val_89
      float[batch,77,768] val_90
      float[batch,77,768] val_91
      float[batch,77,3072] val_92
      float[batch,77,768] val_93
      float[batch,77,768] val_94
      float[batch,77,768] val_95
      float[batch,77,768] val_96
      float[batch,77,768] val_97
      float[batch,77,3072] val_98
      float[batch,77,768] val_99
   >
{
   val_75 = Gather <axis: int = 0> ("text.transformer.embeddings.word_embeddings.weight_fp16", text)
   embedding = Cast <to: int = 1> (val_75)
   add_46 = Add (embedding, embedding_1)
   add_55 = Add (add_46, embedding_2)
   [node_layer_norm] layer_norm = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_55, "text.transformer.embeddings.LayerNorm.weight", "text.transformer.embeddings.LayerNorm.bias")
   val_76 = MatMul (layer_norm, val_3)
   [node_linear] linear = Add (val_76, "text.transformer.encoder.layer.0.attention.self.query.bias")
   val_77 = MatMul (layer_norm, val_4)
   linear_1 = Add (val_77, "text.transformer.encoder.layer.0.attention.self.key.bias")
   val_78 = MatMul (layer_norm, val_5)
   [pad_keep] text_keep = Gather <axis: int = 0> (text_pad_keep, text)
   [val_52_f] val_52_f = Unsqueeze (text_keep, text_row_axes)
   [bitwise_and_1_f] bitwise_and_1_f = Sub (val_52_f, text_one)
   val_52_f_bias = Mul (bitwise_and_1_f, text_scale)
   val_52_f_mask = Add (val_52_f_bias, text_q_axis)
   [node_scaled_dot_product_attention] scaled_dot_product_attention = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, scale: float = 0.125, softcap: float = 0> (linear, linear_1, val_78, val_52_f_mask)
   val_79 = MatMul (scaled_dot_product_attention, val_6)
   linear_3 = Add (val_79, "text.transformer.encoder.layer.0.attention.output.dense.bias")
   add_168 = Add (linear_3, layer_norm)
   layer_norm_1 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_168, "text.transformer.encoder.layer.0.attention.output.LayerNorm.weight", "text.transformer.encoder.layer.0.attention.output.LayerNorm.bias")
   val_80 = MatMul (layer_norm_1, val_7)
   linear_4 = Add (val_80, "text.transformer.encoder.layer.0.intermediate.dense.bias")
   [node_gelu] gelu = Gelu <approximate: string = "none"> (linear_4)
   val_81 = MatMul (gelu, val_8)
   linear_5 = Add (val_81, "text.transformer.encoder.layer.0.output.dense.bias")
   add_193 = Add (linear_5, layer_norm_1)
   layer_norm_2 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_193, "text.transformer.encoder.layer.0.output.LayerNorm.weight", "text.transformer.encoder.layer.0.output.LayerNorm.bias")
   val_82 = MatMul (layer_norm_2, val_9)
   linear_6 = Add (val_82, "text.transformer.encoder.layer.1.attention.self.query.bias")
   val_83 = MatMul (layer_norm_2, val_10)
   linear_7 = Add (val_83, "text.transformer.encoder.layer.1.attention.self.key.bias")
   val_84 = MatMul (layer_norm_2, val_11)
   scaled_dot_product_attention_1 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, scale: float = 0.125, softcap: float = 0> (linear_6, linear_7, val_84, val_52_f_mask)
   val_85 = MatMul (scaled_dot_product_attention_1, val_12)
   linear_9 = Add (val_85, "text.transformer.encoder.layer.1.attention.output.dense.bias")
   add_266 = Add (linear_9, layer_norm_2)
   layer_norm_3 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_266, "text.transformer.encoder.layer.1.attention.output.LayerNorm.weight", "text.transformer.encoder.layer.1.attention.output.LayerNorm.bias")
   val_86 = MatMul (layer_norm_3, val_13)
   linear_10 = Add (val_86, "text.transformer.encoder.layer.1.intermediate.dense.bias")
   gelu_1 = Gelu <approximate: string = "none"> (linear_10)
   val_87 = MatMul (gelu_1, val_14)
   linear_11 = Add (val_87, "text.transformer.encoder.layer.1.output.dense.bias")
   add_291 = Add (linear_11, layer_norm_3)
   layer_norm_4 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_291, "text.transformer.encoder.layer.1.output.LayerNorm.weight", "text.transformer.encoder.layer.1.output.LayerNorm.bias")
   val_88 = MatMul (layer_norm_4, val_15)
   linear_12 = Add (val_88, "text.transformer.encoder.layer.2.attention.self.query.bias")
   val_89 = MatMul (layer_norm_4, val_16)
   linear_13 = Add (val_89, "text.transformer.encoder.layer.2.attention.self.key.bias")
   val_90 = MatMul (layer_norm_4, val_17)
   scaled_dot_product_attention_2 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, scale: float = 0.125, softcap: float = 0> (linear_12, linear_13, val_90, val_52_f_mask)
   val_91 = MatMul (scaled_dot_product_attention_2, val_18)
   linear_15 = Add (val_91, "text.transformer.encoder.layer.2.attention.output.dense.bias")
   add_364 = Add (linear_15, layer_norm_4)
   layer_norm_5 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_364, "text.transformer.encoder.layer.2.attention.output.LayerNorm.weight", "text.transformer.encoder.layer.2.attention.output.LayerNorm.bias")
   val_92 = MatMul (layer_norm_5, val_19)
   linear_16 = Add (val_92, "text.transformer.encoder.layer.2.intermediate.dense.bias")
   gelu_2 = Gelu <approximate: string = "none"> (linear_16)
   val_93 = MatMul (gelu_2, val_20)
   linear_17 = Add (val_93, "text.transformer.encoder.layer.2.output.dense.bias")
   add_389 = Add (linear_17, layer_norm_5)
   layer_norm_6 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_389, "text.transformer.encoder.layer.2.output.LayerNorm.weight", "text.transformer.encoder.layer.2.output.LayerNorm.bias")
   val_94 = MatMul (layer_norm_6, val_21)
   linear_18 = Add (val_94, "text.transformer.encoder.layer.3.attention.self.query.bias")
   val_95 = MatMul (layer_norm_6, val_22)
   linear_19 = Add (val_95, "text.transformer.encoder.layer.3.attention.self.key.bias")
   val_96 = MatMul (layer_norm_6, val_23)
   scaled_dot_product_attention_3 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, scale: float = 0.125, softcap: float = 0> (linear_18, linear_19, val_96, val_52_f_mask)
   val_97 = MatMul (scaled_dot_product_attention_3, val_24)
   linear_21 = Add (val_97, "text.transformer.encoder.layer.3.attention.output.dense.bias")
   add_462 = Add (linear_21, layer_norm_6)
   layer_norm_7 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_462, "text.transformer.encoder.layer.3.attention.output.LayerNorm.weight", "text.transformer.encoder.layer.3.attention.output.LayerNorm.bias")
   val_98 = MatMul (layer_norm_7, val_25)
   linear_22 = Add (val_98, "text.transformer.encoder.layer.3.intermediate.dense.bias")
   gelu_3 = Gelu <approximate: string = "none"> (linear_22)
   val_99 = MatMul (gelu_3, val_26)
   linear_23 = Add (val_99, "text.transformer.encoder.layer.3.output.dense.bias")
   add_487 = Add (linear_23, layer_norm_7)
   layer_norm_8 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_487, "text.transformer.encoder.layer.3.output.LayerNorm.weight", "text.transformer.encoder.layer.3.output.LayerNorm.bias")
   val_100 = MatMul (layer_norm_8, val_27)
   linear_24 = Add (val_100, "text.transformer.encoder.layer.4.attention.self.query.bias")
   val_101 = MatMul (layer_norm_8, val_28)
   linear_25 = Add (val_101, "text.transformer.encoder.layer.4.attention.self.key.bias")
   val_102 = MatMul (layer_norm_8, val_29)
   scaled_dot_product_attention_4 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, scale: float = 0.125, softcap: float = 0> (linear_24, linear_25, val_102, val_52_f_mask)
   val_103 = MatMul (scaled_dot_product_attention_4, val_30)
   linear_27 = Add (val_103, "text.transformer.encoder.layer.4.attention.output.dense.bias")
   add_560 = Add (linear_27, layer_norm_8)
   layer_norm_9 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_560, "text.transformer.encoder.layer.4.attention.output.LayerNorm.weight", "text.transformer.encoder.layer.4.attention.output.LayerNorm.bias")
   val_104 = MatMul (layer_norm_9, val_31)
   linear_28 = Add (val_104, "text.transformer.encoder.layer.4.intermediate.dense.bias")
   gelu_4 = Gelu <approximate: string = "none"> (linear_28)
   val_105 = MatMul (gelu_4, val_32)
   linear_29 = Add (val_105, "text.transformer.encoder.layer.4.output.dense.bias")
   add_585 = Add (linear_29, layer_norm_9)
   layer_norm_10 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_585, "text.transformer.encoder.layer.4.output.LayerNorm.weight", "text.transformer.encoder.layer.4.output.LayerNorm.bias")
   val_106 = MatMul (layer_norm_10, val_33)
   linear_30 = Add (val_106, "text.transformer.encoder.layer.5.attention.self.query.bias")
   val_107 = MatMul (layer_norm_10, val_34)
   linear_31 = Add (val_107, "text.transformer.encoder.layer.5.attention.self.key.bias")
   val_108 = MatMul (layer_norm_10, val_35)
   scaled_dot_product_attention_5 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, scale: float = 0.125, softcap: float = 0> (linear_30, linear_31, val_108, val_52_f_mask)
   val_109 = MatMul (scaled_dot_product_attention_5, val_36)
   linear_33 = Add (val_109, "text.transformer.encoder.layer.5.attention.output.dense.bias")
   add_658 = Add (linear_33, layer_norm_10)
   layer_norm_11 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_658, "text.transformer.encoder.layer.5.attention.output.LayerNorm.weight", "text.transformer.encoder.layer.5.attention.output.LayerNorm.bias")
   val_110 = MatMul (layer_norm_11, val_37)
   linear_34 = Add (val_110, "text.transformer.encoder.layer.5.intermediate.dense.bias")
   gelu_5 = Gelu <approximate: string = "none"> (linear_34)
   val_111 = MatMul (gelu_5, val_38)
   linear_35 = Add (val_111, "text.transformer.encoder.layer.5.output.dense.bias")
   add_683 = Add (linear_35, layer_norm_11)
   layer_norm_12 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_683, "text.transformer.encoder.layer.5.output.LayerNorm.weight", "text.transformer.encoder.layer.5.output.LayerNorm.bias")
   val_112 = MatMul (layer_norm_12, val_39)
   linear_36 = Add (val_112, "text.transformer.encoder.layer.6.attention.self.query.bias")
   val_113 = MatMul (layer_norm_12, val_40)
   linear_37 = Add (val_113, "text.transformer.encoder.layer.6.attention.self.key.bias")
   val_114 = MatMul (layer_norm_12, val_41)
   scaled_dot_product_attention_6 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, scale: float = 0.125, softcap: float = 0> (linear_36, linear_37, val_114, val_52_f_mask)
   val_115 = MatMul (scaled_dot_product_attention_6, val_42)
   linear_39 = Add (val_115, "text.transformer.encoder.layer.6.attention.output.dense.bias")
   add_756 = Add (linear_39, layer_norm_12)
   layer_norm_13 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_756, "text.transformer.encoder.layer.6.attention.output.LayerNorm.weight", "text.transformer.encoder.layer.6.attention.output.LayerNorm.bias")
   val_116 = MatMul (layer_norm_13, val_43)
   linear_40 = Add (val_116, "text.transformer.encoder.layer.6.intermediate.dense.bias")
   gelu_6 = Gelu <approximate: string = "none"> (linear_40)
   val_117 = MatMul (gelu_6, val_44)
   linear_41 = Add (val_117, "text.transformer.encoder.layer.6.output.dense.bias")
   add_781 = Add (linear_41, layer_norm_13)
   layer_norm_14 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_781, "text.transformer.encoder.layer.6.output.LayerNorm.weight", "text.transformer.encoder.layer.6.output.LayerNorm.bias")
   val_118 = MatMul (layer_norm_14, val_45)
   linear_42 = Add (val_118, "text.transformer.encoder.layer.7.attention.self.query.bias")
   val_119 = MatMul (layer_norm_14, val_46)
   linear_43 = Add (val_119, "text.transformer.encoder.layer.7.attention.self.key.bias")
   val_120 = MatMul (layer_norm_14, val_47)
   scaled_dot_product_attention_7 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, scale: float = 0.125, softcap: float = 0> (linear_42, linear_43, val_120, val_52_f_mask)
   val_121 = MatMul (scaled_dot_product_attention_7, val_48)
   linear_45 = Add (val_121, "text.transformer.encoder.layer.7.attention.output.dense.bias")
   add_854 = Add (linear_45, layer_norm_14)
   layer_norm_15 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_854, "text.transformer.encoder.layer.7.attention.output.LayerNorm.weight", "text.transformer.encoder.layer.7.attention.output.LayerNorm.bias")
   val_122 = MatMul (layer_norm_15, val_49)
   linear_46 = Add (val_122, "text.transformer.encoder.layer.7.intermediate.dense.bias")
   gelu_7 = Gelu <approximate: string = "none"> (linear_46)
   val_123 = MatMul (gelu_7, val_50)
   linear_47 = Add (val_123, "text.transformer.encoder.layer.7.output.dense.bias")
   add_879 = Add (linear_47, layer_norm_15)
   layer_norm_16 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_879, "text.transformer.encoder.layer.7.output.LayerNorm.weight", "text.transformer.encoder.layer.7.output.LayerNorm.bias")
   val_124 = MatMul (layer_norm_16, val_51)
   linear_48 = Add (val_124, "text.transformer.encoder.layer.8.attention.self.query.bias")
   val_125 = MatMul (layer_norm_16, val_52)
   linear_49 = Add (val_125, "text.transformer.encoder.layer.8.attention.self.key.bias")
   val_126 = MatMul (layer_norm_16, val_53)
   scaled_dot_product_attention_8 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, scale: float = 0.125, softcap: float = 0> (linear_48, linear_49, val_126, val_52_f_mask)
   val_127 = MatMul (scaled_dot_product_attention_8, val_54)
   linear_51 = Add (val_127, "text.transformer.encoder.layer.8.attention.output.dense.bias")
   add_952 = Add (linear_51, layer_norm_16)
   layer_norm_17 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_952, "text.transformer.encoder.layer.8.attention.output.LayerNorm.weight", "text.transformer.encoder.layer.8.attention.output.LayerNorm.bias")
   val_128 = MatMul (layer_norm_17, val_55)
   linear_52 = Add (val_128, "text.transformer.encoder.layer.8.intermediate.dense.bias")
   gelu_8 = Gelu <approximate: string = "none"> (linear_52)
   val_129 = MatMul (gelu_8, val_56)
   linear_53 = Add (val_129, "text.transformer.encoder.layer.8.output.dense.bias")
   add_977 = Add (linear_53, layer_norm_17)
   layer_norm_18 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_977, "text.transformer.encoder.layer.8.output.LayerNorm.weight", "text.transformer.encoder.layer.8.output.LayerNorm.bias")
   val_130 = MatMul (layer_norm_18, val_57)
   linear_54 = Add (val_130, "text.transformer.encoder.layer.9.attention.self.query.bias")
   val_131 = MatMul (layer_norm_18, val_58)
   linear_55 = Add (val_131, "text.transformer.encoder.layer.9.attention.self.key.bias")
   val_132 = MatMul (layer_norm_18, val_59)
   scaled_dot_product_attention_9 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, scale: float = 0.125, softcap: float = 0> (linear_54, linear_55, val_132, val_52_f_mask)
   val_133 = MatMul (scaled_dot_product_attention_9, val_60)
   linear_57 = Add (val_133, "text.transformer.encoder.layer.9.attention.output.dense.bias")
   add_1050 = Add (linear_57, layer_norm_18)
   layer_norm_19 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_1050, "text.transformer.encoder.layer.9.attention.output.LayerNorm.weight", "text.transformer.encoder.layer.9.attention.output.LayerNorm.bias")
   val_134 = MatMul (layer_norm_19, val_61)
   linear_58 = Add (val_134, "text.transformer.encoder.layer.9.intermediate.dense.bias")
   gelu_9 = Gelu <approximate: string = "none"> (linear_58)
   val_135 = MatMul (gelu_9, val_62)
   linear_59 = Add (val_135, "text.transformer.encoder.layer.9.output.dense.bias")
   add_1075 = Add (linear_59, layer_norm_19)
   layer_norm_20 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_1075, "text.transformer.encoder.layer.9.output.LayerNorm.weight", "text.transformer.encoder.layer.9.output.LayerNorm.bias")
   val_136 = MatMul (layer_norm_20, val_63)
   linear_60 = Add (val_136, "text.transformer.encoder.layer.10.attention.self.query.bias")
   val_137 = MatMul (layer_norm_20, val_64)
   linear_61 = Add (val_137, "text.transformer.encoder.layer.10.attention.self.key.bias")
   val_138 = MatMul (layer_norm_20, val_65)
   scaled_dot_product_attention_10 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, scale: float = 0.125, softcap: float = 0> (linear_60, linear_61, val_138, val_52_f_mask)
   val_139 = MatMul (scaled_dot_product_attention_10, val_66)
   linear_63 = Add (val_139, "text.transformer.encoder.layer.10.attention.output.dense.bias")
   add_1148 = Add (linear_63, layer_norm_20)
   layer_norm_21 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_1148, "text.transformer.encoder.layer.10.attention.output.LayerNorm.weight", "text.transformer.encoder.layer.10.attention.output.LayerNorm.bias")
   val_140 = MatMul (layer_norm_21, val_67)
   linear_64 = Add (val_140, "text.transformer.encoder.layer.10.intermediate.dense.bias")
   gelu_10 = Gelu <approximate: string = "none"> (linear_64)
   val_141 = MatMul (gelu_10, val_68)
   linear_65 = Add (val_141, "text.transformer.encoder.layer.10.output.dense.bias")
   add_1173 = Add (linear_65, layer_norm_21)
   layer_norm_22 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_1173, "text.transformer.encoder.layer.10.output.LayerNorm.weight", "text.transformer.encoder.layer.10.output.LayerNorm.bias")
   val_142 = MatMul (layer_norm_22, val_69)
   linear_66 = Add (val_142, "text.transformer.encoder.layer.11.attention.self.query.bias")
   val_143 = MatMul (layer_norm_22, val_70)
   linear_67 = Add (val_143, "text.transformer.encoder.layer.11.attention.self.key.bias")
   val_144 = MatMul (layer_norm_22, val_71)
   scaled_dot_product_attention_11 = Attention <is_causal: int = 0, kv_num_heads: int = 12, q_num_heads: int = 12, qk_matmul_output_mode: int = 0, scale: float = 0.125, softcap: float = 0> (linear_66, linear_67, val_144, val_52_f_mask)
   val_145 = MatMul (scaled_dot_product_attention_11, val_72)
   linear_69 = Add (val_145, "text.transformer.encoder.layer.11.attention.output.dense.bias")
   add_1246 = Add (linear_69, layer_norm_22)
   layer_norm_23 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_1246, "text.transformer.encoder.layer.11.attention.output.LayerNorm.weight", "text.transformer.encoder.layer.11.attention.output.LayerNorm.bias")
   val_146 = MatMul (layer_norm_23, val_73)
   linear_70 = Add (val_146, "text.transformer.encoder.layer.11.intermediate.dense.bias")
   gelu_11 = Gelu <approximate: string = "none"> (linear_70)
   val_147 = MatMul (gelu_11, val_74)
   linear_71 = Add (val_147, "text.transformer.encoder.layer.11.output.dense.bias")
   add_1271 = Add (linear_71, layer_norm_23)
   layer_norm_24 = LayerNormalization <axis: int = -1, epsilon: float = 1e-05, stash_type: int = 1> (add_1271, "text.transformer.encoder.layer.11.output.LayerNorm.weight", "text.transformer.encoder.layer.11.output.LayerNorm.bias")
   [unsqueeze_12_f] unsqueeze_12_f = Unsqueeze (text_keep, val_0)
   mul_637 = Mul (layer_norm_24, unsqueeze_12_f)
   sum_1 = ReduceSum <keepdims: int = 0, noop_with_empty_axes: int = 0> (mul_637, val_2)
   [sum_2_f] sum_2_f = ReduceSum <keepdims: int = 1, noop_with_empty_axes: int = 0> (text_keep, val_0)
   [node_div] div = Div (sum_1, sum_2_f)
   linear_72 = Gemm <alpha: float = 1, beta: float = 1, transA: int = 0, transB: int = 1> (div, "text.proj.0.weight")
   gelu_12 = Gelu <approximate: string = "none"> (linear_72)
   linear_73 = Gemm <alpha: float = 1, beta: float = 1, transA: int = 0, transB: int = 1> (gelu_12, "text.proj.2.weight")
   [node_linalg_vector_norm] linalg_vector_norm = ReduceL2 <keepdims: int = 1, noop_with_empty_axes: int = 0> (linear_73, val_0)
   [node_clamp_min] clamp_min = Clip (linalg_vector_norm, val_1)
   text_embedding = Div (linear_73, clamp_min)
}

weights:
embedding_1 FLOAT[1,1,768] 8d5002bda41b
embedding_2 FLOAT[1,77,768] 1b573ff43046
text.proj.0.weight FLOAT[640,768] cabe040773ae
text.proj.2.weight FLOAT[512,640] b8c96d7b7594
text.transformer.embeddings.LayerNorm.bias FLOAT[768] 7aa8493f5748
text.transformer.embeddings.LayerNorm.weight FLOAT[768] e3484ddfbe97
text.transformer.embeddings.word_embeddings.weight_fp16 FLOAT16[250002,768] ce7361b2bb3b
text.transformer.encoder.layer.0.attention.output.LayerNorm.bias FLOAT[768] 6c1f00023e78
text.transformer.encoder.layer.0.attention.output.LayerNorm.weight FLOAT[768] 3c49bb0f3c9d
text.transformer.encoder.layer.0.attention.output.dense.bias FLOAT[768] 7137778e0978
text.transformer.encoder.layer.0.attention.self.key.bias FLOAT[768] 4aab87f22564
text.transformer.encoder.layer.0.attention.self.query.bias FLOAT[768] a84445d8480a
text.transformer.encoder.layer.0.intermediate.dense.bias FLOAT[3072] 9da3e3649daf
text.transformer.encoder.layer.0.output.LayerNorm.bias FLOAT[768] cafae2962776
text.transformer.encoder.layer.0.output.LayerNorm.weight FLOAT[768] d11a0fb1b847
text.transformer.encoder.layer.0.output.dense.bias FLOAT[768] 370414783792
text.transformer.encoder.layer.1.attention.output.LayerNorm.bias FLOAT[768] c8f6589bc230
text.transformer.encoder.layer.1.attention.output.LayerNorm.weight FLOAT[768] 98f73b4376fa
text.transformer.encoder.layer.1.attention.output.dense.bias FLOAT[768] 9bc46d7712e9
text.transformer.encoder.layer.1.attention.self.key.bias FLOAT[768] 05b2c6e3fd08
text.transformer.encoder.layer.1.attention.self.query.bias FLOAT[768] ab94ec1480f6
text.transformer.encoder.layer.1.intermediate.dense.bias FLOAT[3072] c0d5ad96864c
text.transformer.encoder.layer.1.output.LayerNorm.bias FLOAT[768] b60f9aa22f29
text.transformer.encoder.layer.1.output.LayerNorm.weight FLOAT[768] b25073b524f7
text.transformer.encoder.layer.1.output.dense.bias FLOAT[768] e8e07c8af046
text.transformer.encoder.layer.10.attention.output.LayerNorm.bias FLOAT[768] 8f89637cf6e1
text.transformer.encoder.layer.10.attention.output.LayerNorm.weight FLOAT[768] 776dc70f43e6
text.transformer.encoder.layer.10.attention.output.dense.bias FLOAT[768] 090e55256bbc
text.transformer.encoder.layer.10.attention.self.key.bias FLOAT[768] afe030c0f3a9
text.transformer.encoder.layer.10.attention.self.query.bias FLOAT[768] 545e40447392
text.transformer.encoder.layer.10.intermediate.dense.bias FLOAT[3072] 5de3ac86f28d
text.transformer.encoder.layer.10.output.LayerNorm.bias FLOAT[768] 71e94af262a5
text.transformer.encoder.layer.10.output.LayerNorm.weight FLOAT[768] 59024590e285
text.transformer.encoder.layer.10.output.dense.bias FLOAT[768] 4d6ad6470a1a
text.transformer.encoder.layer.11.attention.output.LayerNorm.bias FLOAT[768] 65e4dd579910
text.transformer.encoder.layer.11.attention.output.LayerNorm.weight FLOAT[768] bad4ce47ea71
text.transformer.encoder.layer.11.attention.output.dense.bias FLOAT[768] 0d6727468544
text.transformer.encoder.layer.11.attention.self.key.bias FLOAT[768] cb3fb3b1c0a4
text.transformer.encoder.layer.11.attention.self.query.bias FLOAT[768] f08fd1cd3a9e
text.transformer.encoder.layer.11.intermediate.dense.bias FLOAT[3072] c873f405cd66
text.transformer.encoder.layer.11.output.LayerNorm.bias FLOAT[768] f7c8f3a2efcc
text.transformer.encoder.layer.11.output.LayerNorm.weight FLOAT[768] 7d3090c89078
text.transformer.encoder.layer.11.output.dense.bias FLOAT[768] 5590257d4d6b
text.transformer.encoder.layer.2.attention.output.LayerNorm.bias FLOAT[768] ec40c2442a71
text.transformer.encoder.layer.2.attention.output.LayerNorm.weight FLOAT[768] c47c930884df
text.transformer.encoder.layer.2.attention.output.dense.bias FLOAT[768] dd779340e558
text.transformer.encoder.layer.2.attention.self.key.bias FLOAT[768] be955fdd7c85
text.transformer.encoder.layer.2.attention.self.query.bias FLOAT[768] 75cc659162d5
text.transformer.encoder.layer.2.intermediate.dense.bias FLOAT[3072] c0cbdc4f2f85
text.transformer.encoder.layer.2.output.LayerNorm.bias FLOAT[768] 8eb221d9d763
text.transformer.encoder.layer.2.output.LayerNorm.weight FLOAT[768] 6cd1fa2fc5c3
text.transformer.encoder.layer.2.output.dense.bias FLOAT[768] aa361f166bc5
text.transformer.encoder.layer.3.attention.output.LayerNorm.bias FLOAT[768] c83e9bb65bb8
text.transformer.encoder.layer.3.attention.output.LayerNorm.weight FLOAT[768] 69bed4061c8b
text.transformer.encoder.layer.3.attention.output.dense.bias FLOAT[768] aa5bc3f3c9b3
text.transformer.encoder.layer.3.attention.self.key.bias FLOAT[768] 035b3000332e
text.transformer.encoder.layer.3.attention.self.query.bias FLOAT[768] 50e7e1494ec3
text.transformer.encoder.layer.3.intermediate.dense.bias FLOAT[3072] 10882267f05d
text.transformer.encoder.layer.3.output.LayerNorm.bias FLOAT[768] bfceef57a5b3
text.transformer.encoder.layer.3.output.LayerNorm.weight FLOAT[768] f1f0a1c5f2ef
text.transformer.encoder.layer.3.output.dense.bias FLOAT[768] 5a302929a3ec
text.transformer.encoder.layer.4.attention.output.LayerNorm.bias FLOAT[768] 02b207b22480
text.transformer.encoder.layer.4.attention.output.LayerNorm.weight FLOAT[768] 4288594a89a1
text.transformer.encoder.layer.4.attention.output.dense.bias FLOAT[768] 9c79c7980611
text.transformer.encoder.layer.4.attention.self.key.bias FLOAT[768] f929db9f24cc
text.transformer.encoder.layer.4.attention.self.query.bias FLOAT[768] a8de239c2401
text.transformer.encoder.layer.4.intermediate.dense.bias FLOAT[3072] 1d7bc6ec7ab1
text.transformer.encoder.layer.4.output.LayerNorm.bias FLOAT[768] 459cb136d13d
text.transformer.encoder.layer.4.output.LayerNorm.weight FLOAT[768] dac9a63d705c
text.transformer.encoder.layer.4.output.dense.bias FLOAT[768] 811ba744b6c7
text.transformer.encoder.layer.5.attention.output.LayerNorm.bias FLOAT[768] ddd4748dc0ee
text.transformer.encoder.layer.5.attention.output.LayerNorm.weight FLOAT[768] 0c21e80a6b58
text.transformer.encoder.layer.5.attention.output.dense.bias FLOAT[768] 375ab2ee4cde
text.transformer.encoder.layer.5.attention.self.key.bias FLOAT[768] 2b98c15b312b
text.transformer.encoder.layer.5.attention.self.query.bias FLOAT[768] c4b4d9045b9c
text.transformer.encoder.layer.5.intermediate.dense.bias FLOAT[3072] e7aee3d0e16c
text.transformer.encoder.layer.5.output.LayerNorm.bias FLOAT[768] 40fc45127563
text.transformer.encoder.layer.5.output.LayerNorm.weight FLOAT[768] 9880d54b5faa
text.transformer.encoder.layer.5.output.dense.bias FLOAT[768] 6d99a2a9fc14
text.transformer.encoder.layer.6.attention.output.LayerNorm.bias FLOAT[768] 367cbd53d47c
text.transformer.encoder.layer.6.attention.output.LayerNorm.weight FLOAT[768] 2d9785024c40
text.transformer.encoder.layer.6.attention.output.dense.bias FLOAT[768] 023a68354a0e
text.transformer.encoder.layer.6.attention.self.key.bias FLOAT[768] 09a4bac9bc14
text.transformer.encoder.layer.6.attention.self.query.bias FLOAT[768] b68d21bf5ddf
text.transformer.encoder.layer.6.intermediate.dense.bias FLOAT[3072] 600325056fdb
text.transformer.encoder.layer.6.output.LayerNorm.bias FLOAT[768] 87d62b9cf464
text.transformer.encoder.layer.6.output.LayerNorm.weight FLOAT[768] d4616eeea87a
text.transformer.encoder.layer.6.output.dense.bias FLOAT[768] f45170dfd703
text.transformer.encoder.layer.7.attention.output.LayerNorm.bias FLOAT[768] 6ba320f96f66
text.transformer.encoder.layer.7.attention.output.LayerNorm.weight FLOAT[768] fb690f92342c
text.transformer.encoder.layer.7.attention.output.dense.bias FLOAT[768] 9d702df037e5
text.transformer.encoder.layer.7.attention.self.key.bias FLOAT[768] fb822c00902a
text.transformer.encoder.layer.7.attention.self.query.bias FLOAT[768] 0df762875df0
text.transformer.encoder.layer.7.intermediate.dense.bias FLOAT[3072] b3c8411f9ea3
text.transformer.encoder.layer.7.output.LayerNorm.bias FLOAT[768] a5c4cd9518c6
text.transformer.encoder.layer.7.output.LayerNorm.weight FLOAT[768] 131c61d4b01b
text.transformer.encoder.layer.7.output.dense.bias FLOAT[768] 0b397f6b3bae
text.transformer.encoder.layer.8.attention.output.LayerNorm.bias FLOAT[768] da1cbd70f856
text.transformer.encoder.layer.8.attention.output.LayerNorm.weight FLOAT[768] 8e3cbcfd8b56
text.transformer.encoder.layer.8.attention.output.dense.bias FLOAT[768] 3832a331b31b
text.transformer.encoder.layer.8.attention.self.key.bias FLOAT[768] c8fff53f8dcf
text.transformer.encoder.layer.8.attention.self.query.bias FLOAT[768] a4cfddf09554
text.transformer.encoder.layer.8.intermediate.dense.bias FLOAT[3072] 4fe654ece5eb
text.transformer.encoder.layer.8.output.LayerNorm.bias FLOAT[768] 53156abb2cfd
text.transformer.encoder.layer.8.output.LayerNorm.weight FLOAT[768] 4052f9b546c2
text.transformer.encoder.layer.8.output.dense.bias FLOAT[768] 6211f5abc057
text.transformer.encoder.layer.9.attention.output.LayerNorm.bias FLOAT[768] 8174a587744e
text.transformer.encoder.layer.9.attention.output.LayerNorm.weight FLOAT[768] 42d9d92b2ec0
text.transformer.encoder.layer.9.attention.output.dense.bias FLOAT[768] ef25a80da8e3
text.transformer.encoder.layer.9.attention.self.key.bias FLOAT[768] 8f56fc49b517
text.transformer.encoder.layer.9.attention.self.query.bias FLOAT[768] 27efb4f714f0
text.transformer.encoder.layer.9.intermediate.dense.bias FLOAT[3072] 62de7ce02087
text.transformer.encoder.layer.9.output.LayerNorm.bias FLOAT[768] 97a5eb8bcad6
text.transformer.encoder.layer.9.output.LayerNorm.weight FLOAT[768] 7b75e0262bc3
text.transformer.encoder.layer.9.output.dense.bias FLOAT[768] 9024e9a0ce61
text_one FLOAT[] e00e5eb94441
text_pad_keep FLOAT[250002] 49e563d9ce8a
text_q_axis FLOAT[77,1] 9575b2125169
text_row_axes INT64[2] 0c730b69905c
text_scale FLOAT[] e401200e5808
val_0 INT64[1] 12a3ae445661
val_1 FLOAT[] 6708d9be4956
val_10 FLOAT[768,768] a881056d920e
val_11 FLOAT[768,768] 19648d0de9eb
val_12 FLOAT[768,768] fd2f77290d15
val_13 FLOAT[768,3072] dc9e18ebf02a
val_14 FLOAT[3072,768] 21b2102f0346
val_15 FLOAT[768,768] 59e8873c77e4
val_16 FLOAT[768,768] 0eca54d3ed73
val_17 FLOAT[768,768] 0bc637305c62
val_18 FLOAT[768,768] 4b4d14a3f6ea
val_19 FLOAT[768,3072] 23f14d958106
val_2 INT64[1] 7c9fa136d441
val_20 FLOAT[3072,768] 62217f4d9783
val_21 FLOAT[768,768] 70824d0e9301
val_22 FLOAT[768,768] 1cb9d4e51d2c
val_23 FLOAT[768,768] 4043dc77aa35
val_24 FLOAT[768,768] 2b74dccbbef5
val_25 FLOAT[768,3072] 115fe386137b
val_26 FLOAT[3072,768] 668db0c65e17
val_27 FLOAT[768,768] bc66c7cbd9f2
val_28 FLOAT[768,768] dd307b797e1d
val_29 FLOAT[768,768] a51138e2e1ca
val_3 FLOAT[768,768] b366b138bbfc
val_30 FLOAT[768,768] 3994c406b567
val_31 FLOAT[768,3072] 75507184a551
val_32 FLOAT[3072,768] 6f2414180f29
val_33 FLOAT[768,768] 9965dd6e147c
val_34 FLOAT[768,768] 507bcb26c2cd
val_35 FLOAT[768,768] 6be317e85110
val_36 FLOAT[768,768] 6ae2c2e5e683
val_37 FLOAT[768,3072] 02a001ab85c9
val_38 FLOAT[3072,768] dc5ea195208b
val_39 FLOAT[768,768] d8b4a07d0a13
val_4 FLOAT[768,768] 1c7bb0056b53
val_40 FLOAT[768,768] 92a414478730
val_41 FLOAT[768,768] 4c234172c95c
val_42 FLOAT[768,768] 3e20a78d0fa8
val_43 FLOAT[768,3072] 53eaba22626e
val_44 FLOAT[3072,768] 345f9d28486d
val_45 FLOAT[768,768] d6947760864d
val_46 FLOAT[768,768] d4f7ee8edf6d
val_47 FLOAT[768,768] 81b298669224
val_48 FLOAT[768,768] 64d066280455
val_49 FLOAT[768,3072] 9d4838b4be0d
val_5 FLOAT[768,768] 0499d18928bd
val_50 FLOAT[3072,768] 8d67c15fff90
val_51 FLOAT[768,768] 5d823994ecf8
val_52 FLOAT[768,768] f3a7800a14f8
val_53 FLOAT[768,768] 180452512c9a
val_54 FLOAT[768,768] f990d2ceb56a
val_55 FLOAT[768,3072] 18df0e76c7b2
val_56 FLOAT[3072,768] 89a8ea842010
val_57 FLOAT[768,768] 375ef6f06012
val_58 FLOAT[768,768] ebf8d1f71808
val_59 FLOAT[768,768] 14880546f5cd
val_6 FLOAT[768,768] 32e5450083e2
val_60 FLOAT[768,768] 21e4ef59b66d
val_61 FLOAT[768,3072] 833bd4b8a8f2
val_62 FLOAT[3072,768] f497a6e0fec3
val_63 FLOAT[768,768] d8ba4c69d5fa
val_64 FLOAT[768,768] 6d1b73d4ac1c
val_65 FLOAT[768,768] 9016db6af709
val_66 FLOAT[768,768] 88c7672c021b
val_67 FLOAT[768,3072] ab8360e18e6d
val_68 FLOAT[3072,768] 8df7b310f434
val_69 FLOAT[768,768] 021e70290d4d
val_7 FLOAT[768,3072] ee66876ca4ce
val_70 FLOAT[768,768] 17b82a587a81
val_71 FLOAT[768,768] 390cf747b494
val_72 FLOAT[768,768] b7a5a157978d
val_73 FLOAT[768,3072] cc44925c2ab8
val_74 FLOAT[3072,768] 96fcd5afcc67
val_8 FLOAT[3072,768] 629a9ae04512
val_9 FLOAT[768,768] 5c4e6a2832af
