components:
  ar128_cl4096_1_of_4:
    inputs:
      input_ids:
        shape:
        - 1
        - 128
        dtype: int32
    outputs:
      /model/model/embed_tokens/Gather_output_0:
        shape:
        - 1
        - 128
        - 2560
        dtype: uint16
        quantization_parameters:
          scale: 7.197370679789425e-06
          offset: -30800
  ar128_cl4096_2_of_4:
    inputs:
      /model/model/embed_tokens/Gather_output_0:
        shape:
        - 1
        - 128
        - 2560
        dtype: uint16
        quantization_parameters:
          scale: 7.197370679789425e-06
          offset: -30800
      attention_mask:
        shape:
        - 1
        - 1
        - 128
        - 4096
        dtype: uint16
        quantization_parameters:
          scale: 0.0015259021896696422
          offset: -65535
      position_ids_cos:
        shape:
        - 1
        - 1
        - 128
        - 64
        dtype: uint16
        quantization_parameters:
          scale: 3.0518043793392844e-05
          offset: -32768
      position_ids_sin:
        shape:
        - 1
        - 1
        - 128
        - 64
        dtype: uint16
        quantization_parameters:
          scale: 3.0518043793392844e-05
          offset: -32768
      past_key_0_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 2.345790112112451
          offset: -128
      past_value_0_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.004396342855738843
          offset: -128
      past_key_1_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.4745758537232406
          offset: -128
      past_value_1_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.005273671820759773
          offset: -128
      past_key_2_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.20883603394031525
          offset: -128
      past_value_2_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.0066236164420843124
          offset: -128
      past_key_3_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.20998451113700867
          offset: -128
      past_value_3_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.009058302268385887
          offset: -128
      past_key_4_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.5082661185677596
          offset: -128
      past_value_4_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.010039279237389565
          offset: -128
      past_key_5_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.8604878703440269
          offset: -128
      past_value_5_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.011779437772929668
          offset: -128
      past_key_6_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.32763143224040353
          offset: -128
      past_value_6_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.015452229417860508
          offset: -128
      past_key_7_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.2633135938268947
          offset: -128
      past_value_7_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.023205885663628578
          offset: -128
      past_key_8_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.4432932138442993
          offset: -128
      past_value_8_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.025590643286705017
          offset: -128
      past_key_9_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.198994809248316
          offset: -128
      past_value_9_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.02567124180495739
          offset: -128
      past_key_10_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.3400344252586365
          offset: -128
      past_value_10_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.047514863312244415
          offset: -128
      past_key_11_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.30140187421182946
          offset: -128
      past_value_11_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.029060301818246915
          offset: -128
    outputs:
      /model/model/layers.11/Add_1_output_0:
        shape:
        - 1
        - 128
        - 2560
        dtype: uint16
        quantization_parameters:
          scale: 0.2177089144874876
          offset: -14531
      past_key_0_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 2.345790112112451
          offset: -128
      past_value_0_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.004396342855738843
          offset: -128
      past_key_1_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.4745758537232406
          offset: -128
      past_value_1_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.005273671820759773
          offset: -128
      past_key_2_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.20883603394031525
          offset: -128
      past_value_2_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.0066236164420843124
          offset: -128
      past_key_3_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.20998451113700867
          offset: -128
      past_value_3_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.009058302268385887
          offset: -128
      past_key_4_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.5082661185677596
          offset: -128
      past_value_4_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.010039279237389565
          offset: -128
      past_key_5_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.8604878703440269
          offset: -128
      past_value_5_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.011779437772929668
          offset: -128
      past_key_6_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.32763143224040353
          offset: -128
      past_value_6_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.015452229417860508
          offset: -128
      past_key_7_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.2633135938268947
          offset: -128
      past_value_7_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.023205885663628578
          offset: -128
      past_key_8_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.4432932138442993
          offset: -128
      past_value_8_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.025590643286705017
          offset: -128
      past_key_9_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.198994809248316
          offset: -128
      past_value_9_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.02567124180495739
          offset: -128
      past_key_10_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.3400344252586365
          offset: -128
      past_value_10_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.047514863312244415
          offset: -128
      past_key_11_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.30140187421182946
          offset: -128
      past_value_11_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.029060301818246915
          offset: -128
  ar128_cl4096_3_of_4:
    inputs:
      /model/model/layers.11/Add_1_output_0:
        shape:
        - 1
        - 128
        - 2560
        dtype: uint16
        quantization_parameters:
          scale: 0.2177089144874876
          offset: -14531
      attention_mask:
        shape:
        - 1
        - 1
        - 128
        - 4096
        dtype: uint16
        quantization_parameters:
          scale: 0.0015259021896696422
          offset: -65535
      position_ids_cos:
        shape:
        - 1
        - 1
        - 128
        - 64
        dtype: uint16
        quantization_parameters:
          scale: 3.0518043793392844e-05
          offset: -32768
      position_ids_sin:
        shape:
        - 1
        - 1
        - 128
        - 64
        dtype: uint16
        quantization_parameters:
          scale: 3.0518043793392844e-05
          offset: -32768
      past_key_12_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.2443227916955948
          offset: -128
      past_value_12_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.028940375395647183
          offset: -128
      past_key_13_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.2185003776249923
          offset: -128
      past_value_13_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.17545437812805176
          offset: -128
      past_key_14_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.2587217390537262
          offset: -128
      past_value_14_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.09416288137435913
          offset: -128
      past_key_15_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.18699947867806502
          offset: -128
      past_value_15_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.2346538033072404
          offset: -128
      past_key_16_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.2061757892370224
          offset: -128
      past_value_16_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.19758310092715767
          offset: -128
      past_key_17_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.17539973521795799
          offset: -128
      past_value_17_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.32720860838890076
          offset: -128
      past_key_18_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.19263408690925657
          offset: -128
      past_value_18_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.16426491925096887
          offset: -128
      past_key_19_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.21870377659797668
          offset: -128
      past_value_19_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.06083845153568298
          offset: -128
      past_key_20_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.1891358345746994
          offset: -128
      past_value_20_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.05266200378537178
          offset: -128
      past_key_21_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.19000622606652928
          offset: -128
      past_value_21_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.06354268141618864
          offset: -128
      past_key_22_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.22853676974773407
          offset: -128
      past_value_22_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.19043058156967163
          offset: -128
      past_key_23_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.2349507489542323
          offset: -128
      past_value_23_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.06769621090626153
          offset: -128
    outputs:
      /model/model/layers.23/Add_1_output_0:
        shape:
        - 1
        - 128
        - 2560
        dtype: uint16
        quantization_parameters:
          scale: 0.22199481461480505
          offset: -14218
      past_key_12_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.2443227916955948
          offset: -128
      past_value_12_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.028940375395647183
          offset: -128
      past_key_13_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.2185003776249923
          offset: -128
      past_value_13_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.17545437812805176
          offset: -128
      past_key_14_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.2587217390537262
          offset: -128
      past_value_14_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.09416288137435913
          offset: -128
      past_key_15_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.18699947867806502
          offset: -128
      past_value_15_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.2346538033072404
          offset: -128
      past_key_16_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.2061757892370224
          offset: -128
      past_value_16_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.19758310092715767
          offset: -128
      past_key_17_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.17539973521795799
          offset: -128
      past_value_17_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.32720860838890076
          offset: -128
      past_key_18_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.19263408690925657
          offset: -128
      past_value_18_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.16426491925096887
          offset: -128
      past_key_19_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.21870377659797668
          offset: -128
      past_value_19_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.06083845153568298
          offset: -128
      past_key_20_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.1891358345746994
          offset: -128
      past_value_20_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.05266200378537178
          offset: -128
      past_key_21_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.19000622606652928
          offset: -128
      past_value_21_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.06354268141618864
          offset: -128
      past_key_22_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.22853676974773407
          offset: -128
      past_value_22_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.19043058156967163
          offset: -128
      past_key_23_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.2349507489542323
          offset: -128
      past_value_23_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.06769621090626153
          offset: -128
  ar128_cl4096_4_of_4:
    inputs:
      /model/model/layers.23/Add_1_output_0:
        shape:
        - 1
        - 128
        - 2560
        dtype: uint16
        quantization_parameters:
          scale: 0.22199481461480505
          offset: -14218
      attention_mask:
        shape:
        - 1
        - 1
        - 128
        - 4096
        dtype: uint16
        quantization_parameters:
          scale: 0.0015259021896696422
          offset: -65535
      position_ids_cos:
        shape:
        - 1
        - 1
        - 128
        - 64
        dtype: uint16
        quantization_parameters:
          scale: 3.0518043793392844e-05
          offset: -32768
      position_ids_sin:
        shape:
        - 1
        - 1
        - 128
        - 64
        dtype: uint16
        quantization_parameters:
          scale: 3.0518043793392844e-05
          offset: -32768
      past_key_24_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.2168474873219888
          offset: -128
      past_value_24_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.0993499606847763
          offset: -128
      past_key_25_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.16338758170604706
          offset: -128
      past_value_25_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.08785130083560944
          offset: -128
      past_key_26_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.17082346975803375
          offset: -128
      past_value_26_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.09179645448219119
          offset: -128
      past_key_27_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.15799155831336975
          offset: -128
      past_value_27_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.11564768105745316
          offset: -128
      past_key_28_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.16321668850155327
          offset: -128
      past_value_28_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.12118262193334384
          offset: -128
      past_key_29_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.181901752948761
          offset: -128
      past_value_29_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.35103084534171997
          offset: -128
      past_key_30_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.18615905940532684
          offset: -128
      past_value_30_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.29017698764801025
          offset: -128
      past_key_31_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.19037732481956482
          offset: -128
      past_value_31_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.3625663096510519
          offset: -128
      past_key_32_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.15537120789054812
          offset: -128
      past_value_32_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.45803920115072894
          offset: -128
      past_key_33_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.17338245121512827
          offset: -128
      past_value_33_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.9015522754098487
          offset: -128
      past_key_34_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.17443639039993286
          offset: -128
      past_value_34_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.8790855407714844
          offset: -128
      past_key_35_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.35408836603164673
          offset: -128
      past_value_35_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.470033198595047
          offset: -128
    outputs:
      logits:
        shape:
        - 1
        - 128
        - 151936
        dtype: uint16
        quantization_parameters:
          scale: 0.0013899723052105387
          offset: -26872
      past_key_24_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.2168474873219888
          offset: -128
      past_value_24_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.0993499606847763
          offset: -128
      past_key_25_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.16338758170604706
          offset: -128
      past_value_25_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.08785130083560944
          offset: -128
      past_key_26_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.17082346975803375
          offset: -128
      past_value_26_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.09179645448219119
          offset: -128
      past_key_27_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.15799155831336975
          offset: -128
      past_value_27_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.11564768105745316
          offset: -128
      past_key_28_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.16321668850155327
          offset: -128
      past_value_28_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.12118262193334384
          offset: -128
      past_key_29_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.181901752948761
          offset: -128
      past_value_29_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.35103084534171997
          offset: -128
      past_key_30_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.18615905940532684
          offset: -128
      past_value_30_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.29017698764801025
          offset: -128
      past_key_31_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.19037732481956482
          offset: -128
      past_value_31_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.3625663096510519
          offset: -128
      past_key_32_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.15537120789054812
          offset: -128
      past_value_32_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.45803920115072894
          offset: -128
      past_key_33_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.17338245121512827
          offset: -128
      past_value_33_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.9015522754098487
          offset: -128
      past_key_34_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.17443639039993286
          offset: -128
      past_value_34_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.8790855407714844
          offset: -128
      past_key_35_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.35408836603164673
          offset: -128
      past_value_35_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.470033198595047
          offset: -128
  ar1_cl4096_1_of_4:
    inputs:
      input_ids:
        shape:
        - 1
        - 1
        dtype: int32
    outputs:
      /model/model/embed_tokens/Gather_output_0:
        shape:
        - 1
        - 1
        - 2560
        dtype: uint16
        quantization_parameters:
          scale: 7.197370679789425e-06
          offset: -30800
  ar1_cl4096_2_of_4:
    inputs:
      /model/model/embed_tokens/Gather_output_0:
        shape:
        - 1
        - 1
        - 2560
        dtype: uint16
        quantization_parameters:
          scale: 7.197370679789425e-06
          offset: -30800
      attention_mask:
        shape:
        - 1
        - 1
        - 1
        - 4096
        dtype: uint16
        quantization_parameters:
          scale: 0.0015259021896696422
          offset: -65535
      position_ids_cos:
        shape:
        - 1
        - 1
        - 1
        - 64
        dtype: uint16
        quantization_parameters:
          scale: 3.0518043793392844e-05
          offset: -32768
      position_ids_sin:
        shape:
        - 1
        - 1
        - 1
        - 64
        dtype: uint16
        quantization_parameters:
          scale: 3.0518043793392844e-05
          offset: -32768
      past_key_0_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 2.345790112112451
          offset: -128
      past_value_0_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.004396342855738843
          offset: -128
      past_key_1_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.4745758537232406
          offset: -128
      past_value_1_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.005273671820759773
          offset: -128
      past_key_2_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.20883603394031525
          offset: -128
      past_value_2_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.0066236164420843124
          offset: -128
      past_key_3_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.20998451113700867
          offset: -128
      past_value_3_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.009058302268385887
          offset: -128
      past_key_4_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.5082661185677596
          offset: -128
      past_value_4_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.010039279237389565
          offset: -128
      past_key_5_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.8604878703440269
          offset: -128
      past_value_5_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.011779437772929668
          offset: -128
      past_key_6_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.32763143224040353
          offset: -128
      past_value_6_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.015452229417860508
          offset: -128
      past_key_7_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.2633135938268947
          offset: -128
      past_value_7_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.023205885663628578
          offset: -128
      past_key_8_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.4432932138442993
          offset: -128
      past_value_8_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.025590643286705017
          offset: -128
      past_key_9_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.198994809248316
          offset: -128
      past_value_9_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.02567124180495739
          offset: -128
      past_key_10_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.3400344252586365
          offset: -128
      past_value_10_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.047514863312244415
          offset: -128
      past_key_11_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.30140187421182946
          offset: -128
      past_value_11_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.029060301818246915
          offset: -128
    outputs:
      /model/model/layers.11/Add_1_output_0:
        shape:
        - 1
        - 1
        - 2560
        dtype: uint16
        quantization_parameters:
          scale: 0.2177089144874876
          offset: -14531
      past_key_0_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 2.345790112112451
          offset: -128
      past_value_0_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.004396342855738843
          offset: -128
      past_key_1_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.4745758537232406
          offset: -128
      past_value_1_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.005273671820759773
          offset: -128
      past_key_2_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.20883603394031525
          offset: -128
      past_value_2_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.0066236164420843124
          offset: -128
      past_key_3_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.20998451113700867
          offset: -128
      past_value_3_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.009058302268385887
          offset: -128
      past_key_4_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.5082661185677596
          offset: -128
      past_value_4_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.010039279237389565
          offset: -128
      past_key_5_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.8604878703440269
          offset: -128
      past_value_5_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.011779437772929668
          offset: -128
      past_key_6_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.32763143224040353
          offset: -128
      past_value_6_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.015452229417860508
          offset: -128
      past_key_7_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.2633135938268947
          offset: -128
      past_value_7_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.023205885663628578
          offset: -128
      past_key_8_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.4432932138442993
          offset: -128
      past_value_8_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.025590643286705017
          offset: -128
      past_key_9_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.198994809248316
          offset: -128
      past_value_9_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.02567124180495739
          offset: -128
      past_key_10_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.3400344252586365
          offset: -128
      past_value_10_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.047514863312244415
          offset: -128
      past_key_11_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.30140187421182946
          offset: -128
      past_value_11_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.029060301818246915
          offset: -128
  ar1_cl4096_3_of_4:
    inputs:
      /model/model/layers.11/Add_1_output_0:
        shape:
        - 1
        - 1
        - 2560
        dtype: uint16
        quantization_parameters:
          scale: 0.2177089144874876
          offset: -14531
      attention_mask:
        shape:
        - 1
        - 1
        - 1
        - 4096
        dtype: uint16
        quantization_parameters:
          scale: 0.0015259021896696422
          offset: -65535
      position_ids_cos:
        shape:
        - 1
        - 1
        - 1
        - 64
        dtype: uint16
        quantization_parameters:
          scale: 3.0518043793392844e-05
          offset: -32768
      position_ids_sin:
        shape:
        - 1
        - 1
        - 1
        - 64
        dtype: uint16
        quantization_parameters:
          scale: 3.0518043793392844e-05
          offset: -32768
      past_key_12_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.2443227916955948
          offset: -128
      past_value_12_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.028940375395647183
          offset: -128
      past_key_13_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.2185003776249923
          offset: -128
      past_value_13_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.17545437812805176
          offset: -128
      past_key_14_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.2587217390537262
          offset: -128
      past_value_14_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.09416288137435913
          offset: -128
      past_key_15_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.18699947867806502
          offset: -128
      past_value_15_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.2346538033072404
          offset: -128
      past_key_16_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.2061757892370224
          offset: -128
      past_value_16_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.19758310092715767
          offset: -128
      past_key_17_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.17539973521795799
          offset: -128
      past_value_17_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.32720860838890076
          offset: -128
      past_key_18_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.19263408690925657
          offset: -128
      past_value_18_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.16426491925096887
          offset: -128
      past_key_19_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.21870377659797668
          offset: -128
      past_value_19_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.06083845153568298
          offset: -128
      past_key_20_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.1891358345746994
          offset: -128
      past_value_20_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.05266200378537178
          offset: -128
      past_key_21_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.19000622606652928
          offset: -128
      past_value_21_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.06354268141618864
          offset: -128
      past_key_22_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.22853676974773407
          offset: -128
      past_value_22_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.19043058156967163
          offset: -128
      past_key_23_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.2349507489542323
          offset: -128
      past_value_23_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.06769621090626153
          offset: -128
    outputs:
      /model/model/layers.23/Add_1_output_0:
        shape:
        - 1
        - 1
        - 2560
        dtype: uint16
        quantization_parameters:
          scale: 0.22199481461480505
          offset: -14218
      past_key_12_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.2443227916955948
          offset: -128
      past_value_12_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.028940375395647183
          offset: -128
      past_key_13_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.2185003776249923
          offset: -128
      past_value_13_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.17545437812805176
          offset: -128
      past_key_14_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.2587217390537262
          offset: -128
      past_value_14_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.09416288137435913
          offset: -128
      past_key_15_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.18699947867806502
          offset: -128
      past_value_15_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.2346538033072404
          offset: -128
      past_key_16_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.2061757892370224
          offset: -128
      past_value_16_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.19758310092715767
          offset: -128
      past_key_17_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.17539973521795799
          offset: -128
      past_value_17_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.32720860838890076
          offset: -128
      past_key_18_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.19263408690925657
          offset: -128
      past_value_18_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.16426491925096887
          offset: -128
      past_key_19_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.21870377659797668
          offset: -128
      past_value_19_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.06083845153568298
          offset: -128
      past_key_20_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.1891358345746994
          offset: -128
      past_value_20_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.05266200378537178
          offset: -128
      past_key_21_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.19000622606652928
          offset: -128
      past_value_21_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.06354268141618864
          offset: -128
      past_key_22_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.22853676974773407
          offset: -128
      past_value_22_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.19043058156967163
          offset: -128
      past_key_23_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.2349507489542323
          offset: -128
      past_value_23_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.06769621090626153
          offset: -128
  ar1_cl4096_4_of_4:
    inputs:
      /model/model/layers.23/Add_1_output_0:
        shape:
        - 1
        - 1
        - 2560
        dtype: uint16
        quantization_parameters:
          scale: 0.22199481461480505
          offset: -14218
      attention_mask:
        shape:
        - 1
        - 1
        - 1
        - 4096
        dtype: uint16
        quantization_parameters:
          scale: 0.0015259021896696422
          offset: -65535
      position_ids_cos:
        shape:
        - 1
        - 1
        - 1
        - 64
        dtype: uint16
        quantization_parameters:
          scale: 3.0518043793392844e-05
          offset: -32768
      position_ids_sin:
        shape:
        - 1
        - 1
        - 1
        - 64
        dtype: uint16
        quantization_parameters:
          scale: 3.0518043793392844e-05
          offset: -32768
      past_key_24_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.2168474873219888
          offset: -128
      past_value_24_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.0993499606847763
          offset: -128
      past_key_25_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.16338758170604706
          offset: -128
      past_value_25_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.08785130083560944
          offset: -128
      past_key_26_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.17082346975803375
          offset: -128
      past_value_26_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.09179645448219119
          offset: -128
      past_key_27_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.15799155831336975
          offset: -128
      past_value_27_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.11564768105745316
          offset: -128
      past_key_28_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.16321668850155327
          offset: -128
      past_value_28_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.12118262193334384
          offset: -128
      past_key_29_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.181901752948761
          offset: -128
      past_value_29_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.35103084534171997
          offset: -128
      past_key_30_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.18615905940532684
          offset: -128
      past_value_30_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.29017698764801025
          offset: -128
      past_key_31_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.19037732481956482
          offset: -128
      past_value_31_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.3625663096510519
          offset: -128
      past_key_32_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.15537120789054812
          offset: -128
      past_value_32_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.45803920115072894
          offset: -128
      past_key_33_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.17338245121512827
          offset: -128
      past_value_33_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.9015522754098487
          offset: -128
      past_key_34_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.17443639039993286
          offset: -128
      past_value_34_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.8790855407714844
          offset: -128
      past_key_35_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.35408836603164673
          offset: -128
      past_value_35_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.470033198595047
          offset: -128
    outputs:
      logits:
        shape:
        - 1
        - 1
        - 151936
        dtype: uint16
        quantization_parameters:
          scale: 0.0013899723052105387
          offset: -26872
      past_key_24_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.2168474873219888
          offset: -128
      past_value_24_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.0993499606847763
          offset: -128
      past_key_25_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.16338758170604706
          offset: -128
      past_value_25_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.08785130083560944
          offset: -128
      past_key_26_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.17082346975803375
          offset: -128
      past_value_26_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.09179645448219119
          offset: -128
      past_key_27_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.15799155831336975
          offset: -128
      past_value_27_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.11564768105745316
          offset: -128
      past_key_28_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.16321668850155327
          offset: -128
      past_value_28_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.12118262193334384
          offset: -128
      past_key_29_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.181901752948761
          offset: -128
      past_value_29_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.35103084534171997
          offset: -128
      past_key_30_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.18615905940532684
          offset: -128
      past_value_30_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.29017698764801025
          offset: -128
      past_key_31_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.19037732481956482
          offset: -128
      past_value_31_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.3625663096510519
          offset: -128
      past_key_32_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.15537120789054812
          offset: -128
      past_value_32_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.45803920115072894
          offset: -128
      past_key_33_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.17338245121512827
          offset: -128
      past_value_33_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.9015522754098487
          offset: -128
      past_key_34_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.17443639039993286
          offset: -128
      past_value_34_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.8790855407714844
          offset: -128
      past_key_35_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.35408836603164673
          offset: -128
      past_value_35_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.470033198595047
          offset: -128
precision: w4a16
runtime: genie