diff --git "a/vae_encoder/dd/cache/NhwcConv_0-encoderconv_inConv_meta.json" "b/vae_encoder/dd/cache/NhwcConv_0-encoderconv_inConv_meta.json" new file mode 100644--- /dev/null +++ "b/vae_encoder/dd/cache/NhwcConv_0-encoderconv_inConv_meta.json" @@ -0,0 +1,7962 @@ +{ + "dd_meta_major_version": 1, + "dd_meta_minor_version": 4, + "state_table_updates": [], + "op_list": [ + { + "name": "NhwcConv_0-/encoder/conv_in/Conv", + "type": "SDConv", + "in_args": [ + "Transpose_0_out-init_image.out2_3_0" + ], + "const_args": [ + "NhwcConv_0_weight_NHWC" + ], + "out_args": [ + "NhwcConv_0_out-/encoder/conv_in/Conv_output_0.out2_3_0" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "4" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128", + "3", + "3", + "4" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_0", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_0_out-/encoder/conv_in/Conv_output_0.out2_3_0" + ], + "const_args": [ + "GroupNorm_0_wts_6_0_0" + ], + "out_args": [ + "GroupNorm_0.out12_0.out6_0_0" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_1-/encoder/down_blocks.0/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_0.out12_0.out6_0_0" + ], + "const_args": [ + "NhwcConv_1_weight_NHWC" + ], + "out_args": [ + "NhwcConv_1_out-/encoder/down_blocks.0/resnets.0/conv1/Conv_output_0.out2_3_1" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128", + "3", + "3", + "128" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_1", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_1_out-/encoder/down_blocks.0/resnets.0/conv1/Conv_output_0.out2_3_1" + ], + "const_args": [ + "GroupNorm_1_wts_6_0_1" + ], + "out_args": [ + "GroupNorm_1.out12_1.out6_0_1" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_2-/encoder/down_blocks.0/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_1.out12_1.out6_0_1" + ], + "const_args": [ + "NhwcConv_2_weight_NHWC" + ], + "out_args": [ + "NhwcConv_2_out-/encoder/down_blocks.0/resnets.0/conv2/Conv_output_0.out2_3_2" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128", + "3", + "3", + "128" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/encoder/down_blocks.0/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_0_out-/encoder/conv_in/Conv_output_0.out2_3_0", + "NhwcConv_2_out-/encoder/down_blocks.0/resnets.0/conv2/Conv_output_0.out2_3_2" + ], + "const_args": [], + "out_args": [ + "/encoder/down_blocks.0/resnets.0/Add.out_4_1_0" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_2", + "type": "SDGroupNorm", + "in_args": [ + "/encoder/down_blocks.0/resnets.0/Add.out_4_1_0" + ], + "const_args": [ + "GroupNorm_2_wts_6_0_2" + ], + "out_args": [ + "GroupNorm_2.out12_2.out6_0_2" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_3-/encoder/down_blocks.0/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_2.out12_2.out6_0_2" + ], + "const_args": [ + "NhwcConv_3_weight_NHWC" + ], + "out_args": [ + "NhwcConv_3_out-/encoder/down_blocks.0/resnets.1/conv1/Conv_output_0.out2_3_3" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128", + "3", + "3", + "128" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_3", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_3_out-/encoder/down_blocks.0/resnets.1/conv1/Conv_output_0.out2_3_3" + ], + "const_args": [ + "GroupNorm_3_wts_6_0_3" + ], + "out_args": [ + "GroupNorm_3.out12_3.out6_0_3" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_4-/encoder/down_blocks.0/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_3.out12_3.out6_0_3" + ], + "const_args": [ + "NhwcConv_4_weight_NHWC" + ], + "out_args": [ + "NhwcConv_4_out-/encoder/down_blocks.0/resnets.1/conv2/Conv_output_0.out2_3_4" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128", + "3", + "3", + "128" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/encoder/down_blocks.0/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "/encoder/down_blocks.0/resnets.0/Add.out_4_1_0", + "NhwcConv_4_out-/encoder/down_blocks.0/resnets.1/conv2/Conv_output_0.out2_3_4" + ], + "const_args": [], + "out_args": [ + "/encoder/down_blocks.0/resnets.1/Add.out_4_1_1" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_5-/encoder/down_blocks.0/downsamplers.0/conv/Conv", + "type": "SDConv", + "in_args": [ + "/encoder/down_blocks.0/resnets.1/Add.out_4_1_1" + ], + "const_args": [ + "NhwcConv_5_weight_NHWC" + ], + "out_args": [ + "NhwcConv_5_out-/encoder/down_blocks.0/downsamplers.0/conv/Conv_output_0.out2_3_5" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "strides": { + "type": "int", + "value": [ + "2", + "2" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "1024", + "1024", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128", + "3", + "3", + "128" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_4", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_5_out-/encoder/down_blocks.0/downsamplers.0/conv/Conv_output_0.out2_3_5" + ], + "const_args": [ + "GroupNorm_4_wts_6_0_4" + ], + "out_args": [ + "GroupNorm_4.out12_4.out6_0_4" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_6-/encoder/down_blocks.1/resnets.0/conv_shortcut/Conv", + "type": "SDConv", + "in_args": [ + "NhwcConv_5_out-/encoder/down_blocks.0/downsamplers.0/conv/Conv_output_0.out2_3_5" + ], + "const_args": [ + "NhwcConv_6_weight_NHWC" + ], + "out_args": [ + "NhwcConv_6_out-/encoder/down_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out2_3_6" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "1", + "1", + "128" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "NhwcConv_7-/encoder/down_blocks.1/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_4.out12_4.out6_0_4" + ], + "const_args": [ + "NhwcConv_7_weight_NHWC" + ], + "out_args": [ + "NhwcConv_7_out-/encoder/down_blocks.1/resnets.0/conv1/Conv_output_0.out2_3_7" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "3", + "3", + "128" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_5", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_7_out-/encoder/down_blocks.1/resnets.0/conv1/Conv_output_0.out2_3_7" + ], + "const_args": [ + "GroupNorm_5_wts_6_0_5" + ], + "out_args": [ + "GroupNorm_5.out12_5.out6_0_5" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_8-/encoder/down_blocks.1/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_5.out12_5.out6_0_5" + ], + "const_args": [ + "NhwcConv_8_weight_NHWC" + ], + "out_args": [ + "NhwcConv_8_out-/encoder/down_blocks.1/resnets.0/conv2/Conv_output_0.out2_3_8" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "3", + "3", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/encoder/down_blocks.1/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_6_out-/encoder/down_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out2_3_6", + "NhwcConv_8_out-/encoder/down_blocks.1/resnets.0/conv2/Conv_output_0.out2_3_8" + ], + "const_args": [], + "out_args": [ + "/encoder/down_blocks.1/resnets.0/Add.out_4_1_2" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_6", + "type": "SDGroupNorm", + "in_args": [ + "/encoder/down_blocks.1/resnets.0/Add.out_4_1_2" + ], + "const_args": [ + "GroupNorm_6_wts_6_0_6" + ], + "out_args": [ + "GroupNorm_6.out12_6.out6_0_6" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_9-/encoder/down_blocks.1/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_6.out12_6.out6_0_6" + ], + "const_args": [ + "NhwcConv_9_weight_NHWC" + ], + "out_args": [ + "NhwcConv_9_out-/encoder/down_blocks.1/resnets.1/conv1/Conv_output_0.out2_3_9" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "3", + "3", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_7", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_9_out-/encoder/down_blocks.1/resnets.1/conv1/Conv_output_0.out2_3_9" + ], + "const_args": [ + "GroupNorm_7_wts_6_0_7" + ], + "out_args": [ + "GroupNorm_7.out12_7.out6_0_7" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_10-/encoder/down_blocks.1/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_7.out12_7.out6_0_7" + ], + "const_args": [ + "NhwcConv_10_weight_NHWC" + ], + "out_args": [ + "NhwcConv_10_out-/encoder/down_blocks.1/resnets.1/conv2/Conv_output_0.out2_3_10" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "3", + "3", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/encoder/down_blocks.1/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "/encoder/down_blocks.1/resnets.0/Add.out_4_1_2", + "NhwcConv_10_out-/encoder/down_blocks.1/resnets.1/conv2/Conv_output_0.out2_3_10" + ], + "const_args": [], + "out_args": [ + "/encoder/down_blocks.1/resnets.1/Add.out_4_1_3" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_11-/encoder/down_blocks.1/downsamplers.0/conv/Conv", + "type": "SDConv", + "in_args": [ + "/encoder/down_blocks.1/resnets.1/Add.out_4_1_3" + ], + "const_args": [ + "NhwcConv_11_weight_NHWC" + ], + "out_args": [ + "NhwcConv_11_out-/encoder/down_blocks.1/downsamplers.0/conv/Conv_output_0.out2_3_11" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "strides": { + "type": "int", + "value": [ + "2", + "2" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "3", + "3", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_8", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_11_out-/encoder/down_blocks.1/downsamplers.0/conv/Conv_output_0.out2_3_11" + ], + "const_args": [ + "GroupNorm_8_wts_6_0_8" + ], + "out_args": [ + "GroupNorm_8.out12_8.out6_0_8" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_12-/encoder/down_blocks.2/resnets.0/conv_shortcut/Conv", + "type": "SDConv", + "in_args": [ + "NhwcConv_11_out-/encoder/down_blocks.1/downsamplers.0/conv/Conv_output_0.out2_3_11" + ], + "const_args": [ + "NhwcConv_12_weight_NHWC" + ], + "out_args": [ + "NhwcConv_12_out-/encoder/down_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out2_3_12" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "1", + "1", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "NhwcConv_13-/encoder/down_blocks.2/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_8.out12_8.out6_0_8" + ], + "const_args": [ + "NhwcConv_13_weight_NHWC" + ], + "out_args": [ + "NhwcConv_13_out-/encoder/down_blocks.2/resnets.0/conv1/Conv_output_0.out2_3_13" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_9", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_13_out-/encoder/down_blocks.2/resnets.0/conv1/Conv_output_0.out2_3_13" + ], + "const_args": [ + "GroupNorm_9_wts_6_0_9" + ], + "out_args": [ + "GroupNorm_9.out12_9.out6_0_9" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_14-/encoder/down_blocks.2/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_9.out12_9.out6_0_9" + ], + "const_args": [ + "NhwcConv_14_weight_NHWC" + ], + "out_args": [ + "NhwcConv_14_out-/encoder/down_blocks.2/resnets.0/conv2/Conv_output_0.out2_3_14" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/encoder/down_blocks.2/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_12_out-/encoder/down_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out2_3_12", + "NhwcConv_14_out-/encoder/down_blocks.2/resnets.0/conv2/Conv_output_0.out2_3_14" + ], + "const_args": [], + "out_args": [ + "/encoder/down_blocks.2/resnets.0/Add.out_4_1_4" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_10", + "type": "SDGroupNorm", + "in_args": [ + "/encoder/down_blocks.2/resnets.0/Add.out_4_1_4" + ], + "const_args": [ + "GroupNorm_10_wts_6_0_10" + ], + "out_args": [ + "GroupNorm_10.out12_10.out6_0_10" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_15-/encoder/down_blocks.2/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_10.out12_10.out6_0_10" + ], + "const_args": [ + "NhwcConv_15_weight_NHWC" + ], + "out_args": [ + "NhwcConv_15_out-/encoder/down_blocks.2/resnets.1/conv1/Conv_output_0.out2_3_15" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_11", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_15_out-/encoder/down_blocks.2/resnets.1/conv1/Conv_output_0.out2_3_15" + ], + "const_args": [ + "GroupNorm_11_wts_6_0_11" + ], + "out_args": [ + "GroupNorm_11.out12_11.out6_0_11" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_16-/encoder/down_blocks.2/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_11.out12_11.out6_0_11" + ], + "const_args": [ + "NhwcConv_16_weight_NHWC" + ], + "out_args": [ + "NhwcConv_16_out-/encoder/down_blocks.2/resnets.1/conv2/Conv_output_0.out2_3_16" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/encoder/down_blocks.2/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "/encoder/down_blocks.2/resnets.0/Add.out_4_1_4", + "NhwcConv_16_out-/encoder/down_blocks.2/resnets.1/conv2/Conv_output_0.out2_3_16" + ], + "const_args": [], + "out_args": [ + "/encoder/down_blocks.2/resnets.1/Add.out_4_1_5" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_17-/encoder/down_blocks.2/downsamplers.0/conv/Conv", + "type": "SDConv", + "in_args": [ + "/encoder/down_blocks.2/resnets.1/Add.out_4_1_5" + ], + "const_args": [ + "NhwcConv_17_weight_NHWC" + ], + "out_args": [ + "NhwcConv_17_out-/encoder/down_blocks.2/downsamplers.0/conv/Conv_output_0.out2_3_17" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "strides": { + "type": "int", + "value": [ + "2", + "2" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_12", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_17_out-/encoder/down_blocks.2/downsamplers.0/conv/Conv_output_0.out2_3_17" + ], + "const_args": [ + "GroupNorm_12_wts_6_0_12" + ], + "out_args": [ + "GroupNorm_12.out12_12.out6_0_12" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_18-/encoder/down_blocks.3/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_12.out12_12.out6_0_12" + ], + "const_args": [ + "NhwcConv_18_weight_NHWC" + ], + "out_args": [ + "NhwcConv_18_out-/encoder/down_blocks.3/resnets.0/conv1/Conv_output_0.out2_3_18" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_13", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_18_out-/encoder/down_blocks.3/resnets.0/conv1/Conv_output_0.out2_3_18" + ], + "const_args": [ + "GroupNorm_13_wts_6_0_13" + ], + "out_args": [ + "GroupNorm_13.out12_13.out6_0_13" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_19-/encoder/down_blocks.3/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_13.out12_13.out6_0_13" + ], + "const_args": [ + "NhwcConv_19_weight_NHWC" + ], + "out_args": [ + "NhwcConv_19_out-/encoder/down_blocks.3/resnets.0/conv2/Conv_output_0.out2_3_19" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/encoder/down_blocks.3/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_17_out-/encoder/down_blocks.2/downsamplers.0/conv/Conv_output_0.out2_3_17", + "NhwcConv_19_out-/encoder/down_blocks.3/resnets.0/conv2/Conv_output_0.out2_3_19" + ], + "const_args": [], + "out_args": [ + "/encoder/down_blocks.3/resnets.0/Add.out_4_1_6" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_14", + "type": "SDGroupNorm", + "in_args": [ + "/encoder/down_blocks.3/resnets.0/Add.out_4_1_6" + ], + "const_args": [ + "GroupNorm_14_wts_6_0_14" + ], + "out_args": [ + "GroupNorm_14.out12_14.out6_0_14" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_20-/encoder/down_blocks.3/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_14.out12_14.out6_0_14" + ], + "const_args": [ + "NhwcConv_20_weight_NHWC" + ], + "out_args": [ + "NhwcConv_20_out-/encoder/down_blocks.3/resnets.1/conv1/Conv_output_0.out2_3_20" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_15", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_20_out-/encoder/down_blocks.3/resnets.1/conv1/Conv_output_0.out2_3_20" + ], + "const_args": [ + "GroupNorm_15_wts_6_0_15" + ], + "out_args": [ + "GroupNorm_15.out12_15.out6_0_15" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_21-/encoder/down_blocks.3/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_15.out12_15.out6_0_15" + ], + "const_args": [ + "NhwcConv_21_weight_NHWC" + ], + "out_args": [ + "NhwcConv_21_out-/encoder/down_blocks.3/resnets.1/conv2/Conv_output_0.out2_3_21" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/encoder/down_blocks.3/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "/encoder/down_blocks.3/resnets.0/Add.out_4_1_6", + "NhwcConv_21_out-/encoder/down_blocks.3/resnets.1/conv2/Conv_output_0.out2_3_21" + ], + "const_args": [], + "out_args": [ + "/encoder/down_blocks.3/resnets.1/Add.out_4_1_7" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_16", + "type": "SDGroupNorm", + "in_args": [ + "/encoder/down_blocks.3/resnets.1/Add.out_4_1_7" + ], + "const_args": [ + "GroupNorm_16_wts_6_0_16" + ], + "out_args": [ + "GroupNorm_16.out12_16.out6_0_16" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_22-/encoder/mid_block/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_16.out12_16.out6_0_16" + ], + "const_args": [ + "NhwcConv_22_weight_NHWC" + ], + "out_args": [ + "NhwcConv_22_out-/encoder/mid_block/resnets.0/conv1/Conv_output_0.out2_3_22" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_17", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_22_out-/encoder/mid_block/resnets.0/conv1/Conv_output_0.out2_3_22" + ], + "const_args": [ + "GroupNorm_17_wts_6_0_17" + ], + "out_args": [ + "GroupNorm_17.out12_17.out6_0_17" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_23-/encoder/mid_block/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_17.out12_17.out6_0_17" + ], + "const_args": [ + "NhwcConv_23_weight_NHWC" + ], + "out_args": [ + "NhwcConv_23_out-/encoder/mid_block/resnets.0/conv2/Conv_output_0.out2_3_23" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/encoder/mid_block/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "/encoder/down_blocks.3/resnets.1/Add.out_4_1_7", + "NhwcConv_23_out-/encoder/mid_block/resnets.0/conv2/Conv_output_0.out2_3_23" + ], + "const_args": [], + "out_args": [ + "/encoder/mid_block/resnets.0/Add.out_4_1_8" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_18", + "type": "SDGroupNorm", + "in_args": [ + "/encoder/mid_block/resnets.0/Add.out_4_1_8" + ], + "const_args": [ + "GroupNorm_18_wts_6_1_0" + ], + "out_args": [ + "/encoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out6_1_0" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/encoder/mid_block/attentions.0/to_k/MatMul/MatMulAddFusion", + "type": "SDGemm", + "in_args": [ + "/encoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out6_1_0" + ], + "const_args": [ + "onnx::MatMul_890_2_0_0" + ], + "out_args": [ + "/encoder/mid_block/attentions.0/Reshape_2_output_0_3d.out3_3_0" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "1", + "16384", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "16384", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + } + } + }, + { + "name": "/encoder/mid_block/attentions.0/to_q/MatMul/MatMulAddFusion", + "type": "SDGemm", + "in_args": [ + "/encoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out6_1_0" + ], + "const_args": [ + "onnx::MatMul_889_2_0_1" + ], + "out_args": [ + "/encoder/mid_block/attentions.0/Reshape_1_output_0_3d.out3_3_1" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "1", + "16384", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "16384", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + } + } + }, + { + "name": "/encoder/mid_block/attentions.0/to_v/MatMul/MatMulAddFusion", + "type": "SDGemm", + "in_args": [ + "/encoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out6_1_0" + ], + "const_args": [ + "onnx::MatMul_891_2_0_2" + ], + "out_args": [ + "/encoder/mid_block/attentions.0/Reshape_3_output_0_3d.out3_3_2" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "1", + "16384", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "16384", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + } + } + }, + { + "name": "/encoder/mid_block/attentions.0/MatMul/MatMulScaleFusion/_FusedActivationmha_3_1_0", + "type": "SDMHA_VAE", + "in_args": [ + "/encoder/mid_block/attentions.0/Reshape_1_output_0_3d.out3_3_1", + "/encoder/mid_block/attentions.0/Reshape_2_output_0_3d.out3_3_0", + "/encoder/mid_block/attentions.0/Reshape_3_output_0_3d.out3_3_2" + ], + "const_args": [ + "/encoder/mid_block/attentions.0/MatMul/MatMulScaleFusion/_FusedActivationmha_3_1_0_mask.12_0" + ], + "out_args": [ + "/encoder/mid_block/attentions.0/Transpose_5_output_0_3d.out12_0" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "1" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "16384", + "512", + "16384" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "op_version": { + "type": "str", + "value": [ + "v1.0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/encoder/mid_block/attentions.0/to_out.0/MatMul/MatMulAddFusion", + "type": "SDGemm", + "in_args": [ + "/encoder/mid_block/attentions.0/Transpose_5_output_0_3d.out12_0" + ], + "const_args": [ + "onnx::MatMul_899_2_0_3" + ], + "out_args": [ + "/encoder/mid_block/attentions.0/Reshape_5_output_0.nhwc6_0.out3_0_0" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "1", + "16384", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "16384", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + } + } + }, + { + "name": "/encoder/mid_block/attentions.0/Add", + "type": "SDAdd", + "in_args": [ + "/encoder/mid_block/attentions.0/Reshape_5_output_0.nhwc6_0.out3_0_0", + "/encoder/mid_block/resnets.0/Add.out_4_1_8" + ], + "const_args": [], + "out_args": [ + "/encoder/mid_block/attentions.0/Add_output_0_NHWC.out_4_1_9" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_19", + "type": "SDGroupNorm", + "in_args": [ + "/encoder/mid_block/attentions.0/Add_output_0_NHWC.out_4_1_9" + ], + "const_args": [ + "GroupNorm_19_wts_6_0_18" + ], + "out_args": [ + "GroupNorm_19.out12_19.out6_0_18" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_24-/encoder/mid_block/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_19.out12_19.out6_0_18" + ], + "const_args": [ + "NhwcConv_24_weight_NHWC" + ], + "out_args": [ + "NhwcConv_24_out-/encoder/mid_block/resnets.1/conv1/Conv_output_0.out2_3_24" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_20", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_24_out-/encoder/mid_block/resnets.1/conv1/Conv_output_0.out2_3_24" + ], + "const_args": [ + "GroupNorm_20_wts_6_0_19" + ], + "out_args": [ + "GroupNorm_20.out12_20.out6_0_19" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_25-/encoder/mid_block/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_20.out12_20.out6_0_19" + ], + "const_args": [ + "NhwcConv_25_weight_NHWC" + ], + "out_args": [ + "NhwcConv_25_out-/encoder/mid_block/resnets.1/conv2/Conv_output_0.out2_3_25" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/encoder/mid_block/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "/encoder/mid_block/attentions.0/Add_output_0_NHWC.out_4_1_9", + "NhwcConv_25_out-/encoder/mid_block/resnets.1/conv2/Conv_output_0.out2_3_25" + ], + "const_args": [], + "out_args": [ + "/encoder/mid_block/resnets.1/Add_output_0_NHWC.out_4_1_10" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_21", + "type": "SDGroupNorm", + "in_args": [ + "/encoder/mid_block/resnets.1/Add_output_0_NHWC.out_4_1_10" + ], + "const_args": [ + "GroupNorm_21_wts_6_0_20" + ], + "out_args": [ + "GroupNorm_21.out12_21.out6_0_20" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_26-/encoder/conv_out/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_21.out12_21.out6_0_20" + ], + "const_args": [ + "NhwcConv_26_weight_NHWC" + ], + "out_args": [ + "NhwcConv_26_out-/encoder/conv_out/Conv_output_0.out2_3_26" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "8" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "8", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "NhwcConv_27-/quant_conv/Conv", + "type": "SDConv", + "in_args": [ + "NhwcConv_26_out-/encoder/conv_out/Conv_output_0.out2_3_26" + ], + "const_args": [ + "NhwcConv_27_weight_NHWC" + ], + "out_args": [ + "NhwcConv_27_out-/quant_conv/Conv_output_0.out2_3_27" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "8" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "8" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "8", + "1", + "1", + "8" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + } + ], + "fused_tensors": { + "in": { + "buffer_size": 8388608, + "xrt_arg_id": 0, + "packed_tensors": [ + "Transpose_0_out-init_image.out2_3_0" + ] + }, + "out": { + "buffer_size": 262144, + "xrt_arg_id": 1, + "packed_tensors": [ + "NhwcConv_27_out-/quant_conv/Conv_output_0.out2_3_27" + ] + }, + "scratch": { + "buffer_size": 5654183936, + "xrt_arg_id": 2, + "packed_tensors": [ + "NhwcConv_0_out-/encoder/conv_in/Conv_output_0.out2_3_0", + "GroupNorm_0.out12_0.out6_0_0", + "NhwcConv_1_out-/encoder/down_blocks.0/resnets.0/conv1/Conv_output_0.out2_3_1", + "GroupNorm_1.out12_1.out6_0_1", + "NhwcConv_2_out-/encoder/down_blocks.0/resnets.0/conv2/Conv_output_0.out2_3_2", + "/encoder/down_blocks.0/resnets.0/Add.out_4_1_0", + "GroupNorm_2.out12_2.out6_0_2", + "NhwcConv_3_out-/encoder/down_blocks.0/resnets.1/conv1/Conv_output_0.out2_3_3", + "GroupNorm_3.out12_3.out6_0_3", + "NhwcConv_4_out-/encoder/down_blocks.0/resnets.1/conv2/Conv_output_0.out2_3_4", + "/encoder/down_blocks.0/resnets.1/Add.out_4_1_1", + "NhwcConv_5_out-/encoder/down_blocks.0/downsamplers.0/conv/Conv_output_0.out2_3_5", + "GroupNorm_4.out12_4.out6_0_4", + "NhwcConv_6_out-/encoder/down_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out2_3_6", + "NhwcConv_7_out-/encoder/down_blocks.1/resnets.0/conv1/Conv_output_0.out2_3_7", + "GroupNorm_5.out12_5.out6_0_5", + "NhwcConv_8_out-/encoder/down_blocks.1/resnets.0/conv2/Conv_output_0.out2_3_8", + "/encoder/down_blocks.1/resnets.0/Add.out_4_1_2", + "GroupNorm_6.out12_6.out6_0_6", + "NhwcConv_9_out-/encoder/down_blocks.1/resnets.1/conv1/Conv_output_0.out2_3_9", + "GroupNorm_7.out12_7.out6_0_7", + "NhwcConv_10_out-/encoder/down_blocks.1/resnets.1/conv2/Conv_output_0.out2_3_10", + "/encoder/down_blocks.1/resnets.1/Add.out_4_1_3", + "NhwcConv_11_out-/encoder/down_blocks.1/downsamplers.0/conv/Conv_output_0.out2_3_11", + "GroupNorm_8.out12_8.out6_0_8", + "NhwcConv_12_out-/encoder/down_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out2_3_12", + "NhwcConv_13_out-/encoder/down_blocks.2/resnets.0/conv1/Conv_output_0.out2_3_13", + "GroupNorm_9.out12_9.out6_0_9", + "NhwcConv_14_out-/encoder/down_blocks.2/resnets.0/conv2/Conv_output_0.out2_3_14", + "/encoder/down_blocks.2/resnets.0/Add.out_4_1_4", + "GroupNorm_10.out12_10.out6_0_10", + "NhwcConv_15_out-/encoder/down_blocks.2/resnets.1/conv1/Conv_output_0.out2_3_15", + "GroupNorm_11.out12_11.out6_0_11", + "NhwcConv_16_out-/encoder/down_blocks.2/resnets.1/conv2/Conv_output_0.out2_3_16", + "/encoder/down_blocks.2/resnets.1/Add.out_4_1_5", + "NhwcConv_17_out-/encoder/down_blocks.2/downsamplers.0/conv/Conv_output_0.out2_3_17", + "GroupNorm_12.out12_12.out6_0_12", + "NhwcConv_18_out-/encoder/down_blocks.3/resnets.0/conv1/Conv_output_0.out2_3_18", + "GroupNorm_13.out12_13.out6_0_13", + "NhwcConv_19_out-/encoder/down_blocks.3/resnets.0/conv2/Conv_output_0.out2_3_19", + "/encoder/down_blocks.3/resnets.0/Add.out_4_1_6", + "GroupNorm_14.out12_14.out6_0_14", + "NhwcConv_20_out-/encoder/down_blocks.3/resnets.1/conv1/Conv_output_0.out2_3_20", + "GroupNorm_15.out12_15.out6_0_15", + "NhwcConv_21_out-/encoder/down_blocks.3/resnets.1/conv2/Conv_output_0.out2_3_21", + "/encoder/down_blocks.3/resnets.1/Add.out_4_1_7", + "GroupNorm_16.out12_16.out6_0_16", + "NhwcConv_22_out-/encoder/mid_block/resnets.0/conv1/Conv_output_0.out2_3_22", + "GroupNorm_17.out12_17.out6_0_17", + "NhwcConv_23_out-/encoder/mid_block/resnets.0/conv2/Conv_output_0.out2_3_23", + "/encoder/mid_block/resnets.0/Add.out_4_1_8", + "/encoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out6_1_0", + "/encoder/mid_block/attentions.0/Reshape_2_output_0_3d.out3_3_0", + "/encoder/mid_block/attentions.0/Reshape_1_output_0_3d.out3_3_1", + "/encoder/mid_block/attentions.0/Reshape_3_output_0_3d.out3_3_2", + "/encoder/mid_block/attentions.0/Transpose_5_output_0_3d.out12_0", + "/encoder/mid_block/attentions.0/Reshape_5_output_0.nhwc6_0.out3_0_0", + "/encoder/mid_block/attentions.0/Add_output_0_NHWC.out_4_1_9", + "GroupNorm_19.out12_19.out6_0_18", + "NhwcConv_24_out-/encoder/mid_block/resnets.1/conv1/Conv_output_0.out2_3_24", + "GroupNorm_20.out12_20.out6_0_19", + "NhwcConv_25_out-/encoder/mid_block/resnets.1/conv2/Conv_output_0.out2_3_25", + "/encoder/mid_block/resnets.1/Add_output_0_NHWC.out_4_1_10", + "GroupNorm_21.out12_21.out6_0_20", + "NhwcConv_26_out-/encoder/conv_out/Conv_output_0.out2_3_26" + ] + }, + "const": { + "buffer_size": 42147856, + "xrt_arg_id": 3, + "packed_tensors": [ + "NhwcConv_0_weight_NHWC", + "GroupNorm_0_wts_6_0_0", + "NhwcConv_1_weight_NHWC", + "GroupNorm_1_wts_6_0_1", + "NhwcConv_2_weight_NHWC", + "GroupNorm_2_wts_6_0_2", + "NhwcConv_3_weight_NHWC", + "GroupNorm_3_wts_6_0_3", + "NhwcConv_4_weight_NHWC", + "NhwcConv_5_weight_NHWC", + "GroupNorm_4_wts_6_0_4", + "NhwcConv_6_weight_NHWC", + "NhwcConv_7_weight_NHWC", + "GroupNorm_5_wts_6_0_5", + "NhwcConv_8_weight_NHWC", + "GroupNorm_6_wts_6_0_6", + "NhwcConv_9_weight_NHWC", + "GroupNorm_7_wts_6_0_7", + "NhwcConv_10_weight_NHWC", + "NhwcConv_11_weight_NHWC", + "GroupNorm_8_wts_6_0_8", + "NhwcConv_12_weight_NHWC", + "NhwcConv_13_weight_NHWC", + "GroupNorm_9_wts_6_0_9", + "NhwcConv_14_weight_NHWC", + "GroupNorm_10_wts_6_0_10", + "NhwcConv_15_weight_NHWC", + "GroupNorm_11_wts_6_0_11", + "NhwcConv_16_weight_NHWC", + "NhwcConv_17_weight_NHWC", + "GroupNorm_12_wts_6_0_12", + "NhwcConv_18_weight_NHWC", + "GroupNorm_13_wts_6_0_13", + "NhwcConv_19_weight_NHWC", + "GroupNorm_14_wts_6_0_14", + "NhwcConv_20_weight_NHWC", + "GroupNorm_15_wts_6_0_15", + "NhwcConv_21_weight_NHWC", + "GroupNorm_16_wts_6_0_16", + "NhwcConv_22_weight_NHWC", + "GroupNorm_17_wts_6_0_17", + "NhwcConv_23_weight_NHWC", + "GroupNorm_18_wts_6_1_0", + "onnx::MatMul_890_2_0_0", + "onnx::MatMul_889_2_0_1", + "onnx::MatMul_891_2_0_2", + "/encoder/mid_block/attentions.0/MatMul/MatMulScaleFusion/_FusedActivationmha_3_1_0_mask.12_0", + "onnx::MatMul_899_2_0_3", + "GroupNorm_19_wts_6_0_18", + "NhwcConv_24_weight_NHWC", + "GroupNorm_20_wts_6_0_19", + "NhwcConv_25_weight_NHWC", + "GroupNorm_21_wts_6_0_20", + "NhwcConv_26_weight_NHWC", + "NhwcConv_27_weight_NHWC" + ] + }, + "super_instr": { + "buffer_size": 0, + "xrt_arg_id": 4, + "packed_tensors": [] + } + }, + "tensor_map": { + "Transpose_0_out-init_image.out2_3_0": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1024, + 1024, + 4 + ], + "size_in_bytes": 8388608, + "op_tensor_size": 8388608, + "offset": 0 + }, + "NhwcConv_27_out-/quant_conv/Conv_output_0.out2_3_27": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 8 + ], + "size_in_bytes": 262144, + "op_tensor_size": 262144, + "offset": 0 + }, + "NhwcConv_0_out-/encoder/conv_in/Conv_output_0.out2_3_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1024, + 1024, + 128 + ], + "size_in_bytes": 268435456, + "op_tensor_size": 268435456, + "offset": 0 + }, + "GroupNorm_0.out12_0.out6_0_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1024, + 1024, + 128 + ], + "size_in_bytes": 268435456, + "op_tensor_size": 268435456, + "offset": 268435456 + }, + "NhwcConv_1_out-/encoder/down_blocks.0/resnets.0/conv1/Conv_output_0.out2_3_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1024, + 1024, + 128 + ], + "size_in_bytes": 268435456, + "op_tensor_size": 268435456, + "offset": 536870912 + }, + "GroupNorm_1.out12_1.out6_0_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1024, + 1024, + 128 + ], + "size_in_bytes": 268435456, + "op_tensor_size": 268435456, + "offset": 805306368 + }, + "NhwcConv_2_out-/encoder/down_blocks.0/resnets.0/conv2/Conv_output_0.out2_3_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1024, + 1024, + 128 + ], + "size_in_bytes": 268435456, + "op_tensor_size": 268435456, + "offset": 1073741824 + }, + "/encoder/down_blocks.0/resnets.0/Add.out_4_1_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1024, + 1024, + 128 + ], + "size_in_bytes": 268435456, + "op_tensor_size": 268435456, + "offset": 1342177280 + }, + "GroupNorm_2.out12_2.out6_0_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1024, + 1024, + 128 + ], + "size_in_bytes": 268435456, + "op_tensor_size": 268435456, + "offset": 1610612736 + }, + "NhwcConv_3_out-/encoder/down_blocks.0/resnets.1/conv1/Conv_output_0.out2_3_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1024, + 1024, + 128 + ], + "size_in_bytes": 268435456, + "op_tensor_size": 268435456, + "offset": 1879048192 + }, + "GroupNorm_3.out12_3.out6_0_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1024, + 1024, + 128 + ], + "size_in_bytes": 268435456, + "op_tensor_size": 268435456, + "offset": 2147483648 + }, + "NhwcConv_4_out-/encoder/down_blocks.0/resnets.1/conv2/Conv_output_0.out2_3_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1024, + 1024, + 128 + ], + "size_in_bytes": 268435456, + "op_tensor_size": 268435456, + "offset": 2415919104 + }, + "/encoder/down_blocks.0/resnets.1/Add.out_4_1_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1024, + 1024, + 128 + ], + "size_in_bytes": 268435456, + "op_tensor_size": 268435456, + "offset": 2684354560 + }, + "NhwcConv_5_out-/encoder/down_blocks.0/downsamplers.0/conv/Conv_output_0.out2_3_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 128 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 2952790016 + }, + "GroupNorm_4.out12_4.out6_0_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 128 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 3019898880 + }, + "NhwcConv_6_out-/encoder/down_blocks.1/resnets.0/conv_shortcut/Conv_output_0.out2_3_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 256 + ], + "size_in_bytes": 134217728, + "op_tensor_size": 134217728, + "offset": 3087007744 + }, + "NhwcConv_7_out-/encoder/down_blocks.1/resnets.0/conv1/Conv_output_0.out2_3_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 256 + ], + "size_in_bytes": 134217728, + "op_tensor_size": 134217728, + "offset": 3221225472 + }, + "GroupNorm_5.out12_5.out6_0_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 256 + ], + "size_in_bytes": 134217728, + "op_tensor_size": 134217728, + "offset": 3355443200 + }, + "NhwcConv_8_out-/encoder/down_blocks.1/resnets.0/conv2/Conv_output_0.out2_3_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 256 + ], + "size_in_bytes": 134217728, + "op_tensor_size": 134217728, + "offset": 3489660928 + }, + "/encoder/down_blocks.1/resnets.0/Add.out_4_1_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 256 + ], + "size_in_bytes": 134217728, + "op_tensor_size": 134217728, + "offset": 3623878656 + }, + "GroupNorm_6.out12_6.out6_0_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 256 + ], + "size_in_bytes": 134217728, + "op_tensor_size": 134217728, + "offset": 3758096384 + }, + "NhwcConv_9_out-/encoder/down_blocks.1/resnets.1/conv1/Conv_output_0.out2_3_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 256 + ], + "size_in_bytes": 134217728, + "op_tensor_size": 134217728, + "offset": 3892314112 + }, + "GroupNorm_7.out12_7.out6_0_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 256 + ], + "size_in_bytes": 134217728, + "op_tensor_size": 134217728, + "offset": 4026531840 + }, + "NhwcConv_10_out-/encoder/down_blocks.1/resnets.1/conv2/Conv_output_0.out2_3_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 256 + ], + "size_in_bytes": 134217728, + "op_tensor_size": 134217728, + "offset": 4160749568 + }, + "/encoder/down_blocks.1/resnets.1/Add.out_4_1_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 256 + ], + "size_in_bytes": 134217728, + "op_tensor_size": 134217728, + "offset": 4294967296 + }, + "NhwcConv_11_out-/encoder/down_blocks.1/downsamplers.0/conv/Conv_output_0.out2_3_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 256 + ], + "size_in_bytes": 33554432, + "op_tensor_size": 33554432, + "offset": 4429185024 + }, + "GroupNorm_8.out12_8.out6_0_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 256 + ], + "size_in_bytes": 33554432, + "op_tensor_size": 33554432, + "offset": 4462739456 + }, + "NhwcConv_12_out-/encoder/down_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out2_3_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 512 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 4496293888 + }, + "NhwcConv_13_out-/encoder/down_blocks.2/resnets.0/conv1/Conv_output_0.out2_3_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 512 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 4563402752 + }, + "GroupNorm_9.out12_9.out6_0_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 512 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 4630511616 + }, + "NhwcConv_14_out-/encoder/down_blocks.2/resnets.0/conv2/Conv_output_0.out2_3_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 512 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 4697620480 + }, + "/encoder/down_blocks.2/resnets.0/Add.out_4_1_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 512 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 4764729344 + }, + "GroupNorm_10.out12_10.out6_0_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 512 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 4831838208 + }, + "NhwcConv_15_out-/encoder/down_blocks.2/resnets.1/conv1/Conv_output_0.out2_3_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 512 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 4898947072 + }, + "GroupNorm_11.out12_11.out6_0_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 512 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 4966055936 + }, + "NhwcConv_16_out-/encoder/down_blocks.2/resnets.1/conv2/Conv_output_0.out2_3_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 512 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 5033164800 + }, + "/encoder/down_blocks.2/resnets.1/Add.out_4_1_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 512 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 5100273664 + }, + "NhwcConv_17_out-/encoder/down_blocks.2/downsamplers.0/conv/Conv_output_0.out2_3_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5167382528 + }, + "GroupNorm_12.out12_12.out6_0_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5184159744 + }, + "NhwcConv_18_out-/encoder/down_blocks.3/resnets.0/conv1/Conv_output_0.out2_3_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5200936960 + }, + "GroupNorm_13.out12_13.out6_0_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5217714176 + }, + "NhwcConv_19_out-/encoder/down_blocks.3/resnets.0/conv2/Conv_output_0.out2_3_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5234491392 + }, + "/encoder/down_blocks.3/resnets.0/Add.out_4_1_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5251268608 + }, + "GroupNorm_14.out12_14.out6_0_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5268045824 + }, + "NhwcConv_20_out-/encoder/down_blocks.3/resnets.1/conv1/Conv_output_0.out2_3_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5284823040 + }, + "GroupNorm_15.out12_15.out6_0_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5301600256 + }, + "NhwcConv_21_out-/encoder/down_blocks.3/resnets.1/conv2/Conv_output_0.out2_3_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5318377472 + }, + "/encoder/down_blocks.3/resnets.1/Add.out_4_1_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5335154688 + }, + "GroupNorm_16.out12_16.out6_0_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5351931904 + }, + "NhwcConv_22_out-/encoder/mid_block/resnets.0/conv1/Conv_output_0.out2_3_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5368709120 + }, + "GroupNorm_17.out12_17.out6_0_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5385486336 + }, + "NhwcConv_23_out-/encoder/mid_block/resnets.0/conv2/Conv_output_0.out2_3_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5402263552 + }, + "/encoder/mid_block/resnets.0/Add.out_4_1_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5419040768 + }, + "/encoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out6_1_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 16384, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5435817984 + }, + "/encoder/mid_block/attentions.0/Reshape_2_output_0_3d.out3_3_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 16384, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5452595200 + }, + "/encoder/mid_block/attentions.0/Reshape_1_output_0_3d.out3_3_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 16384, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5469372416 + }, + "/encoder/mid_block/attentions.0/Reshape_3_output_0_3d.out3_3_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 16384, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5486149632 + }, + "/encoder/mid_block/attentions.0/Transpose_5_output_0_3d.out12_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 16384, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5502926848 + }, + "/encoder/mid_block/attentions.0/Reshape_5_output_0.nhwc6_0.out3_0_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5519704064 + }, + "/encoder/mid_block/attentions.0/Add_output_0_NHWC.out_4_1_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5536481280 + }, + "GroupNorm_19.out12_19.out6_0_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5553258496 + }, + "NhwcConv_24_out-/encoder/mid_block/resnets.1/conv1/Conv_output_0.out2_3_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5570035712 + }, + "GroupNorm_20.out12_20.out6_0_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5586812928 + }, + "NhwcConv_25_out-/encoder/mid_block/resnets.1/conv2/Conv_output_0.out2_3_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5603590144 + }, + "/encoder/mid_block/resnets.1/Add_output_0_NHWC.out_4_1_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5620367360 + }, + "GroupNorm_21.out12_21.out6_0_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 5637144576 + }, + "NhwcConv_26_out-/encoder/conv_out/Conv_output_0.out2_3_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 8 + ], + "size_in_bytes": 262144, + "op_tensor_size": 262144, + "offset": 5653921792 + }, + "NhwcConv_0_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 12416 + ], + "size_in_bytes": 12416, + "op_tensor_size": 12416, + "offset": 0, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_0.const", + "file_size": 12416 + }, + "GroupNorm_0_wts_6_0_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "offset": 12416, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_1.const", + "file_size": 512 + }, + "NhwcConv_1_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 174080 + ], + "size_in_bytes": 174080, + "op_tensor_size": 174080, + "offset": 12928, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_2.const", + "file_size": 174080 + }, + "GroupNorm_1_wts_6_0_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "offset": 187008, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_3.const", + "file_size": 512 + }, + "NhwcConv_2_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 174080 + ], + "size_in_bytes": 174080, + "op_tensor_size": 174080, + "offset": 187520, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_4.const", + "file_size": 174080 + }, + "GroupNorm_2_wts_6_0_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "offset": 361600, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_5.const", + "file_size": 512 + }, + "NhwcConv_3_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 174080 + ], + "size_in_bytes": 174080, + "op_tensor_size": 174080, + "offset": 362112, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_6.const", + "file_size": 174080 + }, + "GroupNorm_3_wts_6_0_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "offset": 536192, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_7.const", + "file_size": 512 + }, + "NhwcConv_4_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 174080 + ], + "size_in_bytes": 174080, + "op_tensor_size": 174080, + "offset": 536704, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_8.const", + "file_size": 174080 + }, + "NhwcConv_5_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 174080 + ], + "size_in_bytes": 174080, + "op_tensor_size": 174080, + "offset": 710784, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_9.const", + "file_size": 174080 + }, + "GroupNorm_4_wts_6_0_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "offset": 884864, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_10.const", + "file_size": 512 + }, + "NhwcConv_6_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 40960 + ], + "size_in_bytes": 40960, + "op_tensor_size": 40960, + "offset": 885376, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_11.const", + "file_size": 40960 + }, + "NhwcConv_7_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 348160 + ], + "size_in_bytes": 348160, + "op_tensor_size": 348160, + "offset": 926336, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_12.const", + "file_size": 348160 + }, + "GroupNorm_5_wts_6_0_5": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "offset": 1274496, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_13.const", + "file_size": 1024 + }, + "NhwcConv_8_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 729088 + ], + "size_in_bytes": 729088, + "op_tensor_size": 729088, + "offset": 1275520, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_14.const", + "file_size": 729088 + }, + "GroupNorm_6_wts_6_0_6": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "offset": 2004608, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_15.const", + "file_size": 1024 + }, + "NhwcConv_9_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 729088 + ], + "size_in_bytes": 729088, + "op_tensor_size": 729088, + "offset": 2005632, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_16.const", + "file_size": 729088 + }, + "GroupNorm_7_wts_6_0_7": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "offset": 2734720, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_17.const", + "file_size": 1024 + }, + "NhwcConv_10_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 729088 + ], + "size_in_bytes": 729088, + "op_tensor_size": 729088, + "offset": 2735744, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_18.const", + "file_size": 729088 + }, + "NhwcConv_11_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 729088 + ], + "size_in_bytes": 729088, + "op_tensor_size": 729088, + "offset": 3464832, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_19.const", + "file_size": 729088 + }, + "GroupNorm_8_wts_6_0_8": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "offset": 4193920, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_20.const", + "file_size": 1024 + }, + "NhwcConv_12_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 155648 + ], + "size_in_bytes": 155648, + "op_tensor_size": 155648, + "offset": 4194944, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_21.const", + "file_size": 155648 + }, + "NhwcConv_13_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1458176 + ], + "size_in_bytes": 1458176, + "op_tensor_size": 1458176, + "offset": 4350592, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_22.const", + "file_size": 1458176 + }, + "GroupNorm_9_wts_6_0_9": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 5808768, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_23.const", + "file_size": 2048 + }, + "NhwcConv_14_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 5810816, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_24.const", + "file_size": 2916352 + }, + "GroupNorm_10_wts_6_0_10": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 8727168, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_25.const", + "file_size": 2048 + }, + "NhwcConv_15_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 8729216, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_26.const", + "file_size": 2916352 + }, + "GroupNorm_11_wts_6_0_11": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 11645568, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_27.const", + "file_size": 2048 + }, + "NhwcConv_16_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 11647616, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_28.const", + "file_size": 2916352 + }, + "NhwcConv_17_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 14563968, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_29.const", + "file_size": 2916352 + }, + "GroupNorm_12_wts_6_0_12": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 17480320, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_30.const", + "file_size": 2048 + }, + "NhwcConv_18_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 17482368, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_31.const", + "file_size": 2916352 + }, + "GroupNorm_13_wts_6_0_13": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 20398720, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_32.const", + "file_size": 2048 + }, + "NhwcConv_19_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 20400768, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_33.const", + "file_size": 2916352 + }, + "GroupNorm_14_wts_6_0_14": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 23317120, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_34.const", + "file_size": 2048 + }, + "NhwcConv_20_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 23319168, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_35.const", + "file_size": 2916352 + }, + "GroupNorm_15_wts_6_0_15": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 26235520, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_36.const", + "file_size": 2048 + }, + "NhwcConv_21_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 26237568, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_37.const", + "file_size": 2916352 + }, + "GroupNorm_16_wts_6_0_16": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 29153920, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_38.const", + "file_size": 2048 + }, + "NhwcConv_22_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 29155968, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_39.const", + "file_size": 2916352 + }, + "GroupNorm_17_wts_6_0_17": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 32072320, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_40.const", + "file_size": 2048 + }, + "NhwcConv_23_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 32074368, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_41.const", + "file_size": 2916352 + }, + "GroupNorm_18_wts_6_1_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 34990720, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_42.const", + "file_size": 2048 + }, + "onnx::MatMul_890_2_0_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 299008 + ], + "size_in_bytes": 299008, + "op_tensor_size": 299008, + "offset": 34992768, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_43.const", + "file_size": 299008 + }, + "onnx::MatMul_889_2_0_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 299008 + ], + "size_in_bytes": 299008, + "op_tensor_size": 299008, + "offset": 35291776, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_44.const", + "file_size": 299008 + }, + "onnx::MatMul_891_2_0_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 299008 + ], + "size_in_bytes": 299008, + "op_tensor_size": 299008, + "offset": 35590784, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_45.const", + "file_size": 299008 + }, + "/encoder/mid_block/attentions.0/MatMul/MatMulScaleFusion/_FusedActivationmha_3_1_0_mask.12_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 16384 + ], + "size_in_bytes": 32768, + "op_tensor_size": 32768, + "offset": 35889792, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_46.const", + "file_size": 32768 + }, + "onnx::MatMul_899_2_0_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 299008 + ], + "size_in_bytes": 299008, + "op_tensor_size": 299008, + "offset": 35922560, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_47.const", + "file_size": 299008 + }, + "GroupNorm_19_wts_6_0_18": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 36221568, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_48.const", + "file_size": 2048 + }, + "NhwcConv_24_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 36223616, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_49.const", + "file_size": 2916352 + }, + "GroupNorm_20_wts_6_0_19": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 39139968, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_50.const", + "file_size": 2048 + }, + "NhwcConv_25_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 39142016, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_51.const", + "file_size": 2916352 + }, + "GroupNorm_21_wts_6_0_20": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 42058368, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_52.const", + "file_size": 2048 + }, + "NhwcConv_26_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 87040 + ], + "size_in_bytes": 87040, + "op_tensor_size": 87040, + "offset": 42060416, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_53.const", + "file_size": 87040 + }, + "NhwcConv_27_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 400 + ], + "size_in_bytes": 400, + "op_tensor_size": 400, + "offset": 42147456, + "file_name": "cache/NhwcConv_0-encoderconv_inConv_54.const", + "file_size": 400 + } + }, + "aux_info": {} +} \ No newline at end of file