components:
  ar128_cl4096_1_of_3:
    inputs:
      input_ids:
        shape:
        - 1
        - 128
        dtype: int32
    outputs:
      /model/model/embed_tokens/Gather_output_0:
        shape:
        - 1
        - 128
        - 3072
        dtype: uint16
        quantization_parameters:
          scale: 8.687509536888685e-06
          offset: -27540
  ar128_cl4096_2_of_3:
    inputs:
      /model/model/embed_tokens/Gather_output_0:
        shape:
        - 1
        - 128
        - 3072
        dtype: uint16
        quantization_parameters:
          scale: 8.687509536888685e-06
          offset: -27540
      attention_mask:
        shape:
        - 1
        - 1
        - 128
        - 4096
        dtype: uint16
        quantization_parameters:
          scale: 0.0007629510948348211
          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: 0.15415094792842865
          offset: -128
      past_value_0_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.0076334718614816666
          offset: -128
      past_key_1_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.09501257538795471
          offset: -128
      past_value_1_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.018255468428604246
          offset: -128
      past_key_2_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.12992745637893677
          offset: -128
      past_value_2_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.025279987515426997
          offset: -128
      past_key_3_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.1447511507770208
          offset: -128
      past_value_3_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.028148770332336426
          offset: -128
      past_key_4_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.12199506233996293
          offset: -128
      past_value_4_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.02944980831596795
          offset: -128
      past_key_5_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.12595602869987488
          offset: -128
      past_value_5_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.028479063883423805
          offset: -128
      past_key_6_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.12635421752929688
          offset: -128
      past_value_6_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.025955546647310257
          offset: -128
      past_key_7_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.13450774550437927
          offset: -128
      past_value_7_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.024671796738632082
          offset: -128
      past_key_8_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.1284546439103254
          offset: -128
      past_value_8_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.030749136582016945
          offset: -128
      past_key_9_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.16326546669006348
          offset: -128
      past_value_9_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.03575011125699742
          offset: -128
      past_key_10_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.1229240819811821
          offset: -128
      past_value_10_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.032141510397195816
          offset: -128
      past_key_11_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.13410881161689758
          offset: -128
      past_value_11_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.02836606258482445
          offset: -128
      past_key_12_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.11923930794000626
          offset: -128
      past_value_12_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.03199498108991488
          offset: -128
      past_key_13_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.14966769481268455
          offset: -128
      past_value_13_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.028661564579160196
          offset: -128
    outputs:
      /model/model/layers.13/Add_1_output_0:
        shape:
        - 1
        - 128
        - 3072
        dtype: uint16
        quantization_parameters:
          scale: 0.011241219636185578
          offset: -31773
      past_key_0_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.15415094792842865
          offset: -128
      past_value_0_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.0076334718614816666
          offset: -128
      past_key_1_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.09501257538795471
          offset: -128
      past_value_1_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.018255468428604246
          offset: -128
      past_key_2_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.12992745637893677
          offset: -128
      past_value_2_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.025279987515426997
          offset: -128
      past_key_3_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.1447511507770208
          offset: -128
      past_value_3_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.028148770332336426
          offset: -128
      past_key_4_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.12199506233996293
          offset: -128
      past_value_4_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.02944980831596795
          offset: -128
      past_key_5_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.12595602869987488
          offset: -128
      past_value_5_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.028479063883423805
          offset: -128
      past_key_6_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.12635421752929688
          offset: -128
      past_value_6_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.025955546647310257
          offset: -128
      past_key_7_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.13450774550437927
          offset: -128
      past_value_7_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.024671796738632082
          offset: -128
      past_key_8_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.1284546439103254
          offset: -128
      past_value_8_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.030749136582016945
          offset: -128
      past_key_9_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.16326546669006348
          offset: -128
      past_value_9_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.03575011125699742
          offset: -128
      past_key_10_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.1229240819811821
          offset: -128
      past_value_10_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.032141510397195816
          offset: -128
      past_key_11_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.13410881161689758
          offset: -128
      past_value_11_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.02836606258482445
          offset: -128
      past_key_12_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.11923930794000626
          offset: -128
      past_value_12_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.03199498108991488
          offset: -128
      past_key_13_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.14966769481268455
          offset: -128
      past_value_13_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.028661564579160196
          offset: -128
  ar128_cl4096_3_of_3:
    inputs:
      /model/model/layers.13/Add_1_output_0:
        shape:
        - 1
        - 128
        - 3072
        dtype: uint16
        quantization_parameters:
          scale: 0.011241219636185578
          offset: -31773
      attention_mask:
        shape:
        - 1
        - 1
        - 128
        - 4096
        dtype: uint16
        quantization_parameters:
          scale: 0.0007629510948348211
          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_14_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.13996938269908035
          offset: -128
      past_value_14_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.028563514469176765
          offset: -128
      past_key_15_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.125163213474544
          offset: -128
      past_value_15_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.027315943259892502
          offset: -128
      past_key_16_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.12602750522883857
          offset: -128
      past_value_16_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.033199068158864975
          offset: -128
      past_key_17_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.13656628131866455
          offset: -128
      past_value_17_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.030598552208247146
          offset: -128
      past_key_18_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.11889628913458877
          offset: -128
      past_value_18_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.032141249626874924
          offset: -128
      past_key_19_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.13976185023784637
          offset: -128
      past_value_19_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.032557712765190545
          offset: -128
      past_key_20_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.11934278142733837
          offset: -128
      past_value_20_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.027935825288295746
          offset: -128
      past_key_21_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.13167049377922
          offset: -128
      past_value_21_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.04469357803463936
          offset: -128
      past_key_22_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.13328912689929873
          offset: -128
      past_value_22_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.03745593875646591
          offset: -128
      past_key_23_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.12827403151144193
          offset: -128
      past_value_23_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.032266516238451004
          offset: -128
      past_key_24_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.14478352665901184
          offset: -128
      past_value_24_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.04626827540360098
          offset: -128
      past_key_25_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.17984355716254768
          offset: -128
      past_value_25_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.04539092882411686
          offset: -128
      past_key_26_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.1248558149563046
          offset: -128
      past_value_26_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.05314682051539421
          offset: -128
      past_key_27_in:
        shape:
        - 8
        - 1
        - 128
        - 3968
        dtype: uint8
        quantization_parameters:
          scale: 0.13326138257980347
          offset: -128
      past_value_27_in:
        shape:
        - 8
        - 1
        - 3968
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.04699636995792389
          offset: -128
    outputs:
      logits:
        shape:
        - 1
        - 128
        - 128256
        dtype: uint16
        quantization_parameters:
          scale: 0.0006455201492198974
          offset: -24293
      past_key_14_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.13996938269908035
          offset: -128
      past_value_14_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.028563514469176765
          offset: -128
      past_key_15_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.125163213474544
          offset: -128
      past_value_15_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.027315943259892502
          offset: -128
      past_key_16_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.12602750522883857
          offset: -128
      past_value_16_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.033199068158864975
          offset: -128
      past_key_17_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.13656628131866455
          offset: -128
      past_value_17_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.030598552208247146
          offset: -128
      past_key_18_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.11889628913458877
          offset: -128
      past_value_18_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.032141249626874924
          offset: -128
      past_key_19_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.13976185023784637
          offset: -128
      past_value_19_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.032557712765190545
          offset: -128
      past_key_20_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.11934278142733837
          offset: -128
      past_value_20_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.027935825288295746
          offset: -128
      past_key_21_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.13167049377922
          offset: -128
      past_value_21_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.04469357803463936
          offset: -128
      past_key_22_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.13328912689929873
          offset: -128
      past_value_22_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.03745593875646591
          offset: -128
      past_key_23_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.12827403151144193
          offset: -128
      past_value_23_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.032266516238451004
          offset: -128
      past_key_24_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.14478352665901184
          offset: -128
      past_value_24_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.04626827540360098
          offset: -128
      past_key_25_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.17984355716254768
          offset: -128
      past_value_25_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.04539092882411686
          offset: -128
      past_key_26_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.1248558149563046
          offset: -128
      past_value_26_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.05314682051539421
          offset: -128
      past_key_27_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.13326138257980347
          offset: -128
      past_value_27_out:
        shape:
        - 8
        - 1
        - 128
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.04699636995792389
          offset: -128
  ar1_cl4096_1_of_3:
    inputs:
      input_ids:
        shape:
        - 1
        - 1
        dtype: int32
    outputs:
      /model/model/embed_tokens/Gather_output_0:
        shape:
        - 1
        - 1
        - 3072
        dtype: uint16
        quantization_parameters:
          scale: 8.687509536888685e-06
          offset: -27540
  ar1_cl4096_2_of_3:
    inputs:
      /model/model/embed_tokens/Gather_output_0:
        shape:
        - 1
        - 1
        - 3072
        dtype: uint16
        quantization_parameters:
          scale: 8.687509536888685e-06
          offset: -27540
      attention_mask:
        shape:
        - 1
        - 1
        - 1
        - 4096
        dtype: uint16
        quantization_parameters:
          scale: 0.0007629510948348211
          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: 0.15415094792842865
          offset: -128
      past_value_0_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.0076334718614816666
          offset: -128
      past_key_1_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.09501257538795471
          offset: -128
      past_value_1_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.018255468428604246
          offset: -128
      past_key_2_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.12992745637893677
          offset: -128
      past_value_2_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.025279987515426997
          offset: -128
      past_key_3_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.1447511507770208
          offset: -128
      past_value_3_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.028148770332336426
          offset: -128
      past_key_4_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.12199506233996293
          offset: -128
      past_value_4_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.02944980831596795
          offset: -128
      past_key_5_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.12595602869987488
          offset: -128
      past_value_5_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.028479063883423805
          offset: -128
      past_key_6_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.12635421752929688
          offset: -128
      past_value_6_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.025955546647310257
          offset: -128
      past_key_7_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.13450774550437927
          offset: -128
      past_value_7_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.024671796738632082
          offset: -128
      past_key_8_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.1284546439103254
          offset: -128
      past_value_8_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.030749136582016945
          offset: -128
      past_key_9_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.16326546669006348
          offset: -128
      past_value_9_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.03575011125699742
          offset: -128
      past_key_10_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.1229240819811821
          offset: -128
      past_value_10_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.032141510397195816
          offset: -128
      past_key_11_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.13410881161689758
          offset: -128
      past_value_11_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.02836606258482445
          offset: -128
      past_key_12_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.11923930794000626
          offset: -128
      past_value_12_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.03199498108991488
          offset: -128
      past_key_13_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.14966769481268455
          offset: -128
      past_value_13_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.028661564579160196
          offset: -128
    outputs:
      /model/model/layers.13/Add_1_output_0:
        shape:
        - 1
        - 1
        - 3072
        dtype: uint16
        quantization_parameters:
          scale: 0.011241219636185578
          offset: -31773
      past_key_0_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.15415094792842865
          offset: -128
      past_value_0_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.0076334718614816666
          offset: -128
      past_key_1_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.09501257538795471
          offset: -128
      past_value_1_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.018255468428604246
          offset: -128
      past_key_2_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.12992745637893677
          offset: -128
      past_value_2_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.025279987515426997
          offset: -128
      past_key_3_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.1447511507770208
          offset: -128
      past_value_3_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.028148770332336426
          offset: -128
      past_key_4_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.12199506233996293
          offset: -128
      past_value_4_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.02944980831596795
          offset: -128
      past_key_5_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.12595602869987488
          offset: -128
      past_value_5_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.028479063883423805
          offset: -128
      past_key_6_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.12635421752929688
          offset: -128
      past_value_6_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.025955546647310257
          offset: -128
      past_key_7_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.13450774550437927
          offset: -128
      past_value_7_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.024671796738632082
          offset: -128
      past_key_8_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.1284546439103254
          offset: -128
      past_value_8_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.030749136582016945
          offset: -128
      past_key_9_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.16326546669006348
          offset: -128
      past_value_9_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.03575011125699742
          offset: -128
      past_key_10_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.1229240819811821
          offset: -128
      past_value_10_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.032141510397195816
          offset: -128
      past_key_11_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.13410881161689758
          offset: -128
      past_value_11_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.02836606258482445
          offset: -128
      past_key_12_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.11923930794000626
          offset: -128
      past_value_12_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.03199498108991488
          offset: -128
      past_key_13_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.14966769481268455
          offset: -128
      past_value_13_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.028661564579160196
          offset: -128
  ar1_cl4096_3_of_3:
    inputs:
      /model/model/layers.13/Add_1_output_0:
        shape:
        - 1
        - 1
        - 3072
        dtype: uint16
        quantization_parameters:
          scale: 0.011241219636185578
          offset: -31773
      attention_mask:
        shape:
        - 1
        - 1
        - 1
        - 4096
        dtype: uint16
        quantization_parameters:
          scale: 0.0007629510948348211
          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_14_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.13996938269908035
          offset: -128
      past_value_14_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.028563514469176765
          offset: -128
      past_key_15_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.125163213474544
          offset: -128
      past_value_15_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.027315943259892502
          offset: -128
      past_key_16_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.12602750522883857
          offset: -128
      past_value_16_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.033199068158864975
          offset: -128
      past_key_17_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.13656628131866455
          offset: -128
      past_value_17_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.030598552208247146
          offset: -128
      past_key_18_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.11889628913458877
          offset: -128
      past_value_18_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.032141249626874924
          offset: -128
      past_key_19_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.13976185023784637
          offset: -128
      past_value_19_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.032557712765190545
          offset: -128
      past_key_20_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.11934278142733837
          offset: -128
      past_value_20_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.027935825288295746
          offset: -128
      past_key_21_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.13167049377922
          offset: -128
      past_value_21_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.04469357803463936
          offset: -128
      past_key_22_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.13328912689929873
          offset: -128
      past_value_22_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.03745593875646591
          offset: -128
      past_key_23_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.12827403151144193
          offset: -128
      past_value_23_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.032266516238451004
          offset: -128
      past_key_24_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.14478352665901184
          offset: -128
      past_value_24_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.04626827540360098
          offset: -128
      past_key_25_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.17984355716254768
          offset: -128
      past_value_25_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.04539092882411686
          offset: -128
      past_key_26_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.1248558149563046
          offset: -128
      past_value_26_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.05314682051539421
          offset: -128
      past_key_27_in:
        shape:
        - 8
        - 1
        - 128
        - 4095
        dtype: uint8
        quantization_parameters:
          scale: 0.13326138257980347
          offset: -128
      past_value_27_in:
        shape:
        - 8
        - 1
        - 4095
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.04699636995792389
          offset: -128
    outputs:
      logits:
        shape:
        - 1
        - 1
        - 128256
        dtype: uint16
        quantization_parameters:
          scale: 0.0006455201492198974
          offset: -24293
      past_key_14_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.13996938269908035
          offset: -128
      past_value_14_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.028563514469176765
          offset: -128
      past_key_15_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.125163213474544
          offset: -128
      past_value_15_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.027315943259892502
          offset: -128
      past_key_16_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.12602750522883857
          offset: -128
      past_value_16_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.033199068158864975
          offset: -128
      past_key_17_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.13656628131866455
          offset: -128
      past_value_17_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.030598552208247146
          offset: -128
      past_key_18_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.11889628913458877
          offset: -128
      past_value_18_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.032141249626874924
          offset: -128
      past_key_19_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.13976185023784637
          offset: -128
      past_value_19_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.032557712765190545
          offset: -128
      past_key_20_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.11934278142733837
          offset: -128
      past_value_20_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.027935825288295746
          offset: -128
      past_key_21_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.13167049377922
          offset: -128
      past_value_21_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.04469357803463936
          offset: -128
      past_key_22_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.13328912689929873
          offset: -128
      past_value_22_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.03745593875646591
          offset: -128
      past_key_23_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.12827403151144193
          offset: -128
      past_value_23_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.032266516238451004
          offset: -128
      past_key_24_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.14478352665901184
          offset: -128
      past_value_24_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.04626827540360098
          offset: -128
      past_key_25_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.17984355716254768
          offset: -128
      past_value_25_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.04539092882411686
          offset: -128
      past_key_26_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.1248558149563046
          offset: -128
      past_value_26_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.05314682051539421
          offset: -128
      past_key_27_out:
        shape:
        - 8
        - 1
        - 128
        - 1
        dtype: uint8
        quantization_parameters:
          scale: 0.13326138257980347
          offset: -128
      past_value_27_out:
        shape:
        - 8
        - 1
        - 1
        - 128
        dtype: uint8
        quantization_parameters:
          scale: 0.04699636995792389
          offset: -128
precision: w4a16
runtime: genie