diff --git "a/common/vae_decoder/dynamic/dd/cache/NhwcConv_0-decoderconv_inConv_meta.json" "b/common/vae_decoder/dynamic/dd/cache/NhwcConv_0-decoderconv_inConv_meta.json" new file mode 100644--- /dev/null +++ "b/common/vae_decoder/dynamic/dd/cache/NhwcConv_0-decoderconv_inConv_meta.json" @@ -0,0 +1,11109 @@ +{ + "dd_meta_major_version": 1, + "dd_meta_minor_version": 4, + "state_table_updates": [], + "op_list": [ + { + "name": "NhwcConv_0-/decoder/conv_in/Conv", + "type": "SDConv", + "in_args": [ + "Transpose_0_out-latent_sample.out2_3_0" + ], + "const_args": [ + "NhwcConv_0_weight_NHWC" + ], + "out_args": [ + "NhwcConv_0_out-/decoder/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": "str", + "value": [ + "1", + "height", + "width", + "16" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_0", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_0_out-/decoder/conv_in/Conv_output_0.out2_3_0" + ], + "const_args": [ + "GroupNorm_0_wts_10_0_0" + ], + "out_args": [ + "GroupNorm_0.out13_0.out10_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": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "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_1-/decoder/mid_block/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_0.out13_0.out10_0_0" + ], + "const_args": [ + "NhwcConv_1_weight_NHWC" + ], + "out_args": [ + "NhwcConv_1_out-/decoder/mid_block/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": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_1", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_1_out-/decoder/mid_block/resnets.0/conv1/Conv_output_0.out2_3_1" + ], + "const_args": [ + "GroupNorm_1_wts_10_0_1" + ], + "out_args": [ + "GroupNorm_1.out13_1.out10_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": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "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_2-/decoder/mid_block/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_1.out13_1.out10_0_1" + ], + "const_args": [ + "NhwcConv_2_weight_NHWC" + ], + "out_args": [ + "NhwcConv_2_out-/decoder/mid_block/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": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/decoder/mid_block/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_0_out-/decoder/conv_in/Conv_output_0.out2_3_0", + "NhwcConv_2_out-/decoder/mid_block/resnets.0/conv2/Conv_output_0.out2_3_2" + ], + "const_args": [], + "out_args": [ + "/decoder/mid_block/resnets.0/Add.out_7_1_0" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "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_2", + "type": "SDGroupNorm", + "in_args": [ + "/decoder/mid_block/resnets.0/Add.out_7_1_0" + ], + "const_args": [ + "GroupNorm_2_wts_10_1_0" + ], + "out_args": [ + "/decoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out10_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": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "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": "/decoder/mid_block/attentions.0/to_k/MatMul", + "type": "SDGemm", + "in_args": [ + "/decoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out10_1_0" + ], + "const_args": [ + "onnx::MatMul_914" + ], + "out_args": [ + "/decoder/mid_block/attentions.0/to_k/Add_output_0.out3_4_0" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "1", + "height*width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height*width", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/decoder/mid_block/attentions.0/to_q/MatMul", + "type": "SDGemm", + "in_args": [ + "/decoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out10_1_0" + ], + "const_args": [ + "onnx::MatMul_913" + ], + "out_args": [ + "/decoder/mid_block/attentions.0/to_q/Add_output_0.out3_4_1" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "1", + "height*width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height*width", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/decoder/mid_block/attentions.0/to_v/MatMul", + "type": "SDGemm", + "in_args": [ + "/decoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out10_1_0" + ], + "const_args": [ + "onnx::MatMul_915" + ], + "out_args": [ + "/decoder/mid_block/attentions.0/to_v/Add_output_0.out3_4_2" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "1", + "height*width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height*width", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/decoder/mid_block/attentions.0/MatMulmha_4_0_0", + "type": "SDMHA_VAE", + "in_args": [ + "/decoder/mid_block/attentions.0/to_q/Add_output_0.out3_4_1", + "/decoder/mid_block/attentions.0/to_k/Add_output_0.out3_4_0", + "/decoder/mid_block/attentions.0/to_v/Add_output_0.out3_4_2" + ], + "const_args": [ + "/decoder/mid_block/attentions.0/MatMulmha_4_0_0_mask.9_0" + ], + "out_args": [ + "/decoder/mid_block/attentions.0/Reshape_4_output_0.out9_0" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "1" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "1", + "height*width", + "512", + "height*width" + ] + }, + "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": "/decoder/mid_block/attentions.0/to_out.0/MatMul", + "type": "SDGemm", + "in_args": [ + "/decoder/mid_block/attentions.0/Reshape_4_output_0.out9_0" + ], + "const_args": [ + "onnx::MatMul_923" + ], + "out_args": [ + "/decoder/mid_block/attentions.0/Transpose_6_output_0_4d.out3_1_0" + ], + "attrs": { + "input_shape": { + "type": "str", + "value": [ + "1", + "height*width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height*width", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/decoder/mid_block/attentions.0/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/mid_block/attentions.0/Transpose_6_output_0_4d.out3_1_0", + "/decoder/mid_block/resnets.0/Add.out_7_1_0" + ], + "const_args": [], + "out_args": [ + "/decoder/mid_block/attentions.0/Add_output_0_NHWC.out_7_1_1" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "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_3", + "type": "SDGroupNorm", + "in_args": [ + "/decoder/mid_block/attentions.0/Add_output_0_NHWC.out_7_1_1" + ], + "const_args": [ + "GroupNorm_3_wts_10_0_2" + ], + "out_args": [ + "GroupNorm_3.out13_3.out10_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": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "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_3-/decoder/mid_block/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_3.out13_3.out10_0_2" + ], + "const_args": [ + "NhwcConv_3_weight_NHWC" + ], + "out_args": [ + "NhwcConv_3_out-/decoder/mid_block/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": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_4", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_3_out-/decoder/mid_block/resnets.1/conv1/Conv_output_0.out2_3_3" + ], + "const_args": [ + "GroupNorm_4_wts_10_0_3" + ], + "out_args": [ + "GroupNorm_4.out13_4.out10_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": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "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_4-/decoder/mid_block/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_4.out13_4.out10_0_3" + ], + "const_args": [ + "NhwcConv_4_weight_NHWC" + ], + "out_args": [ + "NhwcConv_4_out-/decoder/mid_block/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": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/decoder/mid_block/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/mid_block/attentions.0/Add_output_0_NHWC.out_7_1_1", + "NhwcConv_4_out-/decoder/mid_block/resnets.1/conv2/Conv_output_0.out2_3_4" + ], + "const_args": [], + "out_args": [ + "/decoder/mid_block/resnets.1/Add_output_0_NHWC.out_7_1_2" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "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_5", + "type": "SDGroupNorm", + "in_args": [ + "/decoder/mid_block/resnets.1/Add_output_0_NHWC.out_7_1_2" + ], + "const_args": [ + "GroupNorm_5_wts_10_0_4" + ], + "out_args": [ + "GroupNorm_5.out13_5.out10_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": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "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_5-/decoder/up_blocks.0/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_5.out13_5.out10_0_4" + ], + "const_args": [ + "NhwcConv_5_weight_NHWC" + ], + "out_args": [ + "NhwcConv_5_out-/decoder/up_blocks.0/resnets.0/conv1/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" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_6", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_5_out-/decoder/up_blocks.0/resnets.0/conv1/Conv_output_0.out2_3_5" + ], + "const_args": [ + "GroupNorm_6_wts_10_0_5" + ], + "out_args": [ + "GroupNorm_6.out13_6.out10_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": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "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_6-/decoder/up_blocks.0/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_6.out13_6.out10_0_5" + ], + "const_args": [ + "NhwcConv_6_weight_NHWC" + ], + "out_args": [ + "NhwcConv_6_out-/decoder/up_blocks.0/resnets.0/conv2/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": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/decoder/up_blocks.0/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/mid_block/resnets.1/Add_output_0_NHWC.out_7_1_2", + "NhwcConv_6_out-/decoder/up_blocks.0/resnets.0/conv2/Conv_output_0.out2_3_6" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.0/resnets.0/Add_output_0_NHWC.out_7_1_3" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "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_7", + "type": "SDGroupNorm", + "in_args": [ + "/decoder/up_blocks.0/resnets.0/Add_output_0_NHWC.out_7_1_3" + ], + "const_args": [ + "GroupNorm_7_wts_10_0_6" + ], + "out_args": [ + "GroupNorm_7.out13_7.out10_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": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "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_7-/decoder/up_blocks.0/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_7.out13_7.out10_0_6" + ], + "const_args": [ + "NhwcConv_7_weight_NHWC" + ], + "out_args": [ + "NhwcConv_7_out-/decoder/up_blocks.0/resnets.1/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": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_8", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_7_out-/decoder/up_blocks.0/resnets.1/conv1/Conv_output_0.out2_3_7" + ], + "const_args": [ + "GroupNorm_8_wts_10_0_7" + ], + "out_args": [ + "GroupNorm_8.out13_8.out10_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": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "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_8-/decoder/up_blocks.0/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_8.out13_8.out10_0_7" + ], + "const_args": [ + "NhwcConv_8_weight_NHWC" + ], + "out_args": [ + "NhwcConv_8_out-/decoder/up_blocks.0/resnets.1/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": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/decoder/up_blocks.0/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/up_blocks.0/resnets.0/Add_output_0_NHWC.out_7_1_3", + "NhwcConv_8_out-/decoder/up_blocks.0/resnets.1/conv2/Conv_output_0.out2_3_8" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.0/resnets.1/Add_output_0_NHWC.out_7_1_4" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "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_9", + "type": "SDGroupNorm", + "in_args": [ + "/decoder/up_blocks.0/resnets.1/Add_output_0_NHWC.out_7_1_4" + ], + "const_args": [ + "GroupNorm_9_wts_10_0_8" + ], + "out_args": [ + "GroupNorm_9.out13_9.out10_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": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "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_9-/decoder/up_blocks.0/resnets.2/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_9.out13_9.out10_0_8" + ], + "const_args": [ + "NhwcConv_9_weight_NHWC" + ], + "out_args": [ + "NhwcConv_9_out-/decoder/up_blocks.0/resnets.2/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": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_10", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_9_out-/decoder/up_blocks.0/resnets.2/conv1/Conv_output_0.out2_3_9" + ], + "const_args": [ + "GroupNorm_10_wts_10_0_9" + ], + "out_args": [ + "GroupNorm_10.out13_10.out10_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": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "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_10-/decoder/up_blocks.0/resnets.2/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_10.out13_10.out10_0_9" + ], + "const_args": [ + "NhwcConv_10_weight_NHWC" + ], + "out_args": [ + "NhwcConv_10_out-/decoder/up_blocks.0/resnets.2/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": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/decoder/up_blocks.0/resnets.2/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/up_blocks.0/resnets.1/Add_output_0_NHWC.out_7_1_4", + "NhwcConv_10_out-/decoder/up_blocks.0/resnets.2/conv2/Conv_output_0.out2_3_10" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.0/resnets.2/Add_output_0.nhwc6_0.out_7_1_5" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "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": "/decoder/up_blocks.0/upsamplers.0/Resize", + "type": "SDResize", + "in_args": [ + "/decoder/up_blocks.0/resnets.2/Add_output_0.nhwc6_0.out_7_1_5" + ], + "const_args": [ + "/decoder/up_blocks.0/upsamplers.0/Resize.weights6_0" + ], + "out_args": [ + "/decoder/up_blocks.0/upsamplers.0/Resize_output_0.nhwc6_0.out_6_0" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "1", + "height", + "width", + "512" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_11-/decoder/up_blocks.0/upsamplers.0/conv/Conv", + "type": "SDConv", + "in_args": [ + "/decoder/up_blocks.0/upsamplers.0/Resize_output_0.nhwc6_0.out_6_0" + ], + "const_args": [ + "NhwcConv_11_weight_NHWC" + ], + "out_args": [ + "NhwcConv_11_out-/decoder/up_blocks.0/upsamplers.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" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_11", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_11_out-/decoder/up_blocks.0/upsamplers.0/conv/Conv_output_0.out2_3_11" + ], + "const_args": [ + "GroupNorm_11_wts_10_0_10" + ], + "out_args": [ + "GroupNorm_11.out13_11.out10_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": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "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_12-/decoder/up_blocks.1/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_11.out13_11.out10_0_10" + ], + "const_args": [ + "NhwcConv_12_weight_NHWC" + ], + "out_args": [ + "NhwcConv_12_out-/decoder/up_blocks.1/resnets.0/conv1/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": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_12", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_12_out-/decoder/up_blocks.1/resnets.0/conv1/Conv_output_0.out2_3_12" + ], + "const_args": [ + "GroupNorm_12_wts_10_0_11" + ], + "out_args": [ + "GroupNorm_12.out13_12.out10_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": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "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_13-/decoder/up_blocks.1/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_12.out13_12.out10_0_11" + ], + "const_args": [ + "NhwcConv_13_weight_NHWC" + ], + "out_args": [ + "NhwcConv_13_out-/decoder/up_blocks.1/resnets.0/conv2/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": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/decoder/up_blocks.1/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_11_out-/decoder/up_blocks.0/upsamplers.0/conv/Conv_output_0.out2_3_11", + "NhwcConv_13_out-/decoder/up_blocks.1/resnets.0/conv2/Conv_output_0.out2_3_13" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.1/resnets.0/Add.out_7_1_6" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "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_13", + "type": "SDGroupNorm", + "in_args": [ + "/decoder/up_blocks.1/resnets.0/Add.out_7_1_6" + ], + "const_args": [ + "GroupNorm_13_wts_10_0_12" + ], + "out_args": [ + "GroupNorm_13.out13_13.out10_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": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "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-/decoder/up_blocks.1/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_13.out13_13.out10_0_12" + ], + "const_args": [ + "NhwcConv_14_weight_NHWC" + ], + "out_args": [ + "NhwcConv_14_out-/decoder/up_blocks.1/resnets.1/conv1/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": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_14", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_14_out-/decoder/up_blocks.1/resnets.1/conv1/Conv_output_0.out2_3_14" + ], + "const_args": [ + "GroupNorm_14_wts_10_0_13" + ], + "out_args": [ + "GroupNorm_14.out13_14.out10_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": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "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-/decoder/up_blocks.1/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_14.out13_14.out10_0_13" + ], + "const_args": [ + "NhwcConv_15_weight_NHWC" + ], + "out_args": [ + "NhwcConv_15_out-/decoder/up_blocks.1/resnets.1/conv2/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": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/decoder/up_blocks.1/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/up_blocks.1/resnets.0/Add.out_7_1_6", + "NhwcConv_15_out-/decoder/up_blocks.1/resnets.1/conv2/Conv_output_0.out2_3_15" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.1/resnets.1/Add.out_7_1_7" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "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_15", + "type": "SDGroupNorm", + "in_args": [ + "/decoder/up_blocks.1/resnets.1/Add.out_7_1_7" + ], + "const_args": [ + "GroupNorm_15_wts_10_0_14" + ], + "out_args": [ + "GroupNorm_15.out13_15.out10_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": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "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-/decoder/up_blocks.1/resnets.2/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_15.out13_15.out10_0_14" + ], + "const_args": [ + "NhwcConv_16_weight_NHWC" + ], + "out_args": [ + "NhwcConv_16_out-/decoder/up_blocks.1/resnets.2/conv1/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": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_16", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_16_out-/decoder/up_blocks.1/resnets.2/conv1/Conv_output_0.out2_3_16" + ], + "const_args": [ + "GroupNorm_16_wts_10_0_15" + ], + "out_args": [ + "GroupNorm_16.out13_16.out10_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": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "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_17-/decoder/up_blocks.1/resnets.2/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_16.out13_16.out10_0_15" + ], + "const_args": [ + "NhwcConv_17_weight_NHWC" + ], + "out_args": [ + "NhwcConv_17_out-/decoder/up_blocks.1/resnets.2/conv2/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" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/decoder/up_blocks.1/resnets.2/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/up_blocks.1/resnets.1/Add.out_7_1_7", + "NhwcConv_17_out-/decoder/up_blocks.1/resnets.2/conv2/Conv_output_0.out2_3_17" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.1/resnets.2/Add.out_7_1_8" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "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": "/decoder/up_blocks.1/upsamplers.0/Resize", + "type": "SDResize", + "in_args": [ + "/decoder/up_blocks.1/resnets.2/Add.out_7_1_8" + ], + "const_args": [ + "/decoder/up_blocks.1/upsamplers.0/Resize.weights6_1" + ], + "out_args": [ + "/decoder/up_blocks.1/upsamplers.0/Resize_output_0.nhwc6_1.out_6_1" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*height)", + "floor(2.0*width)", + "512" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_18-/decoder/up_blocks.1/upsamplers.0/conv/Conv", + "type": "SDConv", + "in_args": [ + "/decoder/up_blocks.1/upsamplers.0/Resize_output_0.nhwc6_1.out_6_1" + ], + "const_args": [ + "NhwcConv_18_weight_NHWC" + ], + "out_args": [ + "NhwcConv_18_out-/decoder/up_blocks.1/upsamplers.0/conv/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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_17", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_18_out-/decoder/up_blocks.1/upsamplers.0/conv/Conv_output_0.out2_3_18" + ], + "const_args": [ + "GroupNorm_17_wts_10_0_16" + ], + "out_args": [ + "GroupNorm_17.out13_17.out10_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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "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-/decoder/up_blocks.2/resnets.0/conv_shortcut/Conv", + "type": "SDConv", + "in_args": [ + "NhwcConv_18_out-/decoder/up_blocks.1/upsamplers.0/conv/Conv_output_0.out2_3_18" + ], + "const_args": [ + "NhwcConv_19_weight_NHWC" + ], + "out_args": [ + "NhwcConv_19_out-/decoder/up_blocks.2/resnets.0/conv_shortcut/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": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "1", + "1", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_20-/decoder/up_blocks.2/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_17.out13_17.out10_0_16" + ], + "const_args": [ + "NhwcConv_20_weight_NHWC" + ], + "out_args": [ + "NhwcConv_20_out-/decoder/up_blocks.2/resnets.0/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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "512" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_18", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_20_out-/decoder/up_blocks.2/resnets.0/conv1/Conv_output_0.out2_3_20" + ], + "const_args": [ + "GroupNorm_18_wts_10_0_17" + ], + "out_args": [ + "GroupNorm_18.out13_18.out10_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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "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_21-/decoder/up_blocks.2/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_18.out13_18.out10_0_17" + ], + "const_args": [ + "NhwcConv_21_weight_NHWC" + ], + "out_args": [ + "NhwcConv_21_out-/decoder/up_blocks.2/resnets.0/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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "3", + "3", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/decoder/up_blocks.2/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_19_out-/decoder/up_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out2_3_19", + "NhwcConv_21_out-/decoder/up_blocks.2/resnets.0/conv2/Conv_output_0.out2_3_21" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.2/resnets.0/Add.out_7_1_9" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "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_19", + "type": "SDGroupNorm", + "in_args": [ + "/decoder/up_blocks.2/resnets.0/Add.out_7_1_9" + ], + "const_args": [ + "GroupNorm_19_wts_10_0_18" + ], + "out_args": [ + "GroupNorm_19.out13_19.out10_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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "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_22-/decoder/up_blocks.2/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_19.out13_19.out10_0_18" + ], + "const_args": [ + "NhwcConv_22_weight_NHWC" + ], + "out_args": [ + "NhwcConv_22_out-/decoder/up_blocks.2/resnets.1/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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "3", + "3", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_20", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_22_out-/decoder/up_blocks.2/resnets.1/conv1/Conv_output_0.out2_3_22" + ], + "const_args": [ + "GroupNorm_20_wts_10_0_19" + ], + "out_args": [ + "GroupNorm_20.out13_20.out10_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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "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_23-/decoder/up_blocks.2/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_20.out13_20.out10_0_19" + ], + "const_args": [ + "NhwcConv_23_weight_NHWC" + ], + "out_args": [ + "NhwcConv_23_out-/decoder/up_blocks.2/resnets.1/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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "3", + "3", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/decoder/up_blocks.2/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/up_blocks.2/resnets.0/Add.out_7_1_9", + "NhwcConv_23_out-/decoder/up_blocks.2/resnets.1/conv2/Conv_output_0.out2_3_23" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.2/resnets.1/Add.out_7_1_10" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "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_21", + "type": "SDGroupNorm", + "in_args": [ + "/decoder/up_blocks.2/resnets.1/Add.out_7_1_10" + ], + "const_args": [ + "GroupNorm_21_wts_10_0_20" + ], + "out_args": [ + "GroupNorm_21.out13_21.out10_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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "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_24-/decoder/up_blocks.2/resnets.2/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_21.out13_21.out10_0_20" + ], + "const_args": [ + "NhwcConv_24_weight_NHWC" + ], + "out_args": [ + "NhwcConv_24_out-/decoder/up_blocks.2/resnets.2/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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "3", + "3", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_22", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_24_out-/decoder/up_blocks.2/resnets.2/conv1/Conv_output_0.out2_3_24" + ], + "const_args": [ + "GroupNorm_22_wts_10_0_21" + ], + "out_args": [ + "GroupNorm_22.out13_22.out10_0_21" + ], + "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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "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_25-/decoder/up_blocks.2/resnets.2/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_22.out13_22.out10_0_21" + ], + "const_args": [ + "NhwcConv_25_weight_NHWC" + ], + "out_args": [ + "NhwcConv_25_out-/decoder/up_blocks.2/resnets.2/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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "3", + "3", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/decoder/up_blocks.2/resnets.2/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/up_blocks.2/resnets.1/Add.out_7_1_10", + "NhwcConv_25_out-/decoder/up_blocks.2/resnets.2/conv2/Conv_output_0.out2_3_25" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.2/resnets.2/Add.out_7_1_11" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "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": "/decoder/up_blocks.2/upsamplers.0/Resize", + "type": "SDResize", + "in_args": [ + "/decoder/up_blocks.2/resnets.2/Add.out_7_1_11" + ], + "const_args": [ + "/decoder/up_blocks.2/upsamplers.0/Resize.weights6_2" + ], + "out_args": [ + "/decoder/up_blocks.2/upsamplers.0/Resize_output_0.nhwc6_2.out_6_2" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "256" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "256" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_26-/decoder/up_blocks.2/upsamplers.0/conv/Conv", + "type": "SDConv", + "in_args": [ + "/decoder/up_blocks.2/upsamplers.0/Resize_output_0.nhwc6_2.out_6_2" + ], + "const_args": [ + "NhwcConv_26_weight_NHWC" + ], + "out_args": [ + "NhwcConv_26_out-/decoder/up_blocks.2/upsamplers.0/conv/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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "256" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "3", + "3", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_23", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_26_out-/decoder/up_blocks.2/upsamplers.0/conv/Conv_output_0.out2_3_26" + ], + "const_args": [ + "GroupNorm_23_wts_10_0_22" + ], + "out_args": [ + "GroupNorm_23.out13_23.out10_0_22" + ], + "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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "256" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "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_27-/decoder/up_blocks.3/resnets.0/conv_shortcut/Conv", + "type": "SDConv", + "in_args": [ + "NhwcConv_26_out-/decoder/up_blocks.2/upsamplers.0/conv/Conv_output_0.out2_3_26" + ], + "const_args": [ + "NhwcConv_27_weight_NHWC" + ], + "out_args": [ + "NhwcConv_27_out-/decoder/up_blocks.3/resnets.0/conv_shortcut/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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "256" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128", + "1", + "1", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_28-/decoder/up_blocks.3/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_23.out13_23.out10_0_22" + ], + "const_args": [ + "NhwcConv_28_weight_NHWC" + ], + "out_args": [ + "NhwcConv_28_out-/decoder/up_blocks.3/resnets.0/conv1/Conv_output_0.out2_3_28" + ], + "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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "256" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128", + "3", + "3", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_24", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_28_out-/decoder/up_blocks.3/resnets.0/conv1/Conv_output_0.out2_3_28" + ], + "const_args": [ + "GroupNorm_24_wts_10_0_23" + ], + "out_args": [ + "GroupNorm_24.out13_24.out10_0_23" + ], + "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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "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_29-/decoder/up_blocks.3/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_24.out13_24.out10_0_23" + ], + "const_args": [ + "NhwcConv_29_weight_NHWC" + ], + "out_args": [ + "NhwcConv_29_out-/decoder/up_blocks.3/resnets.0/conv2/Conv_output_0.out2_3_29" + ], + "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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128", + "3", + "3", + "128" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/decoder/up_blocks.3/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_27_out-/decoder/up_blocks.3/resnets.0/conv_shortcut/Conv_output_0.out2_3_27", + "NhwcConv_29_out-/decoder/up_blocks.3/resnets.0/conv2/Conv_output_0.out2_3_29" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.3/resnets.0/Add.out_7_1_12" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "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_25", + "type": "SDGroupNorm", + "in_args": [ + "/decoder/up_blocks.3/resnets.0/Add.out_7_1_12" + ], + "const_args": [ + "GroupNorm_25_wts_10_0_24" + ], + "out_args": [ + "GroupNorm_25.out13_25.out10_0_24" + ], + "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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "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_30-/decoder/up_blocks.3/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_25.out13_25.out10_0_24" + ], + "const_args": [ + "NhwcConv_30_weight_NHWC" + ], + "out_args": [ + "NhwcConv_30_out-/decoder/up_blocks.3/resnets.1/conv1/Conv_output_0.out2_3_30" + ], + "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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128", + "3", + "3", + "128" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_26", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_30_out-/decoder/up_blocks.3/resnets.1/conv1/Conv_output_0.out2_3_30" + ], + "const_args": [ + "GroupNorm_26_wts_10_0_25" + ], + "out_args": [ + "GroupNorm_26.out13_26.out10_0_25" + ], + "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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "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_31-/decoder/up_blocks.3/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_26.out13_26.out10_0_25" + ], + "const_args": [ + "NhwcConv_31_weight_NHWC" + ], + "out_args": [ + "NhwcConv_31_out-/decoder/up_blocks.3/resnets.1/conv2/Conv_output_0.out2_3_31" + ], + "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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128", + "3", + "3", + "128" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/decoder/up_blocks.3/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/up_blocks.3/resnets.0/Add.out_7_1_12", + "NhwcConv_31_out-/decoder/up_blocks.3/resnets.1/conv2/Conv_output_0.out2_3_31" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.3/resnets.1/Add.out_7_1_13" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "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_27", + "type": "SDGroupNorm", + "in_args": [ + "/decoder/up_blocks.3/resnets.1/Add.out_7_1_13" + ], + "const_args": [ + "GroupNorm_27_wts_10_0_26" + ], + "out_args": [ + "GroupNorm_27.out13_27.out10_0_26" + ], + "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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "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_32-/decoder/up_blocks.3/resnets.2/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_27.out13_27.out10_0_26" + ], + "const_args": [ + "NhwcConv_32_weight_NHWC" + ], + "out_args": [ + "NhwcConv_32_out-/decoder/up_blocks.3/resnets.2/conv1/Conv_output_0.out2_3_32" + ], + "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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128", + "3", + "3", + "128" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_28", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_32_out-/decoder/up_blocks.3/resnets.2/conv1/Conv_output_0.out2_3_32" + ], + "const_args": [ + "GroupNorm_28_wts_10_0_27" + ], + "out_args": [ + "GroupNorm_28.out13_28.out10_0_27" + ], + "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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "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_33-/decoder/up_blocks.3/resnets.2/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_28.out13_28.out10_0_27" + ], + "const_args": [ + "NhwcConv_33_weight_NHWC" + ], + "out_args": [ + "NhwcConv_33_out-/decoder/up_blocks.3/resnets.2/conv2/Conv_output_0.out2_3_33" + ], + "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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128", + "3", + "3", + "128" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/decoder/up_blocks.3/resnets.2/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/up_blocks.3/resnets.1/Add.out_7_1_13", + "NhwcConv_33_out-/decoder/up_blocks.3/resnets.2/conv2/Conv_output_0.out2_3_33" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.3/resnets.2/Add.out_7_1_14" + ], + "attrs": { + "a_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "b_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "c_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "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_29", + "type": "SDGroupNorm", + "in_args": [ + "/decoder/up_blocks.3/resnets.2/Add.out_7_1_14" + ], + "const_args": [ + "GroupNorm_29_wts_10_0_28" + ], + "out_args": [ + "GroupNorm_29.out13_29.out10_0_28" + ], + "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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "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_34-/decoder/conv_out/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_29.out13_29.out10_0_28" + ], + "const_args": [ + "NhwcConv_34_weight_NHWC" + ], + "out_args": [ + "NhwcConv_34_out-sample.out2_3_34" + ], + "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": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "128" + ] + }, + "output_shape": { + "type": "str", + "value": [ + "1", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "3" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "3", + "3", + "3", + "128" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + } + ], + "fused_tensors": { + "in": { + "buffer_size": 32, + "xrt_arg_id": 0, + "packed_tensors": [ + "Transpose_0_out-latent_sample.out2_3_0" + ] + }, + "out": { + "buffer_size": 8, + "xrt_arg_id": 1, + "packed_tensors": [ + "NhwcConv_34_out-sample.out2_3_34" + ] + }, + "scratch": { + "buffer_size": 67584, + "xrt_arg_id": 2, + "packed_tensors": [ + "NhwcConv_0_out-/decoder/conv_in/Conv_output_0.out2_3_0", + "GroupNorm_0.out13_0.out10_0_0", + "NhwcConv_1_out-/decoder/mid_block/resnets.0/conv1/Conv_output_0.out2_3_1", + "GroupNorm_1.out13_1.out10_0_1", + "NhwcConv_2_out-/decoder/mid_block/resnets.0/conv2/Conv_output_0.out2_3_2", + "/decoder/mid_block/resnets.0/Add.out_7_1_0", + "/decoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out10_1_0", + "/decoder/mid_block/attentions.0/to_k/Add_output_0.out3_4_0", + "/decoder/mid_block/attentions.0/to_q/Add_output_0.out3_4_1", + "/decoder/mid_block/attentions.0/to_v/Add_output_0.out3_4_2", + "/decoder/mid_block/attentions.0/Reshape_4_output_0.out9_0", + "/decoder/mid_block/attentions.0/Transpose_6_output_0_4d.out3_1_0", + "/decoder/mid_block/attentions.0/Add_output_0_NHWC.out_7_1_1", + "GroupNorm_3.out13_3.out10_0_2", + "NhwcConv_3_out-/decoder/mid_block/resnets.1/conv1/Conv_output_0.out2_3_3", + "GroupNorm_4.out13_4.out10_0_3", + "NhwcConv_4_out-/decoder/mid_block/resnets.1/conv2/Conv_output_0.out2_3_4", + "/decoder/mid_block/resnets.1/Add_output_0_NHWC.out_7_1_2", + "GroupNorm_5.out13_5.out10_0_4", + "NhwcConv_5_out-/decoder/up_blocks.0/resnets.0/conv1/Conv_output_0.out2_3_5", + "GroupNorm_6.out13_6.out10_0_5", + "NhwcConv_6_out-/decoder/up_blocks.0/resnets.0/conv2/Conv_output_0.out2_3_6", + "/decoder/up_blocks.0/resnets.0/Add_output_0_NHWC.out_7_1_3", + "GroupNorm_7.out13_7.out10_0_6", + "NhwcConv_7_out-/decoder/up_blocks.0/resnets.1/conv1/Conv_output_0.out2_3_7", + "GroupNorm_8.out13_8.out10_0_7", + "NhwcConv_8_out-/decoder/up_blocks.0/resnets.1/conv2/Conv_output_0.out2_3_8", + "/decoder/up_blocks.0/resnets.1/Add_output_0_NHWC.out_7_1_4", + "GroupNorm_9.out13_9.out10_0_8", + "NhwcConv_9_out-/decoder/up_blocks.0/resnets.2/conv1/Conv_output_0.out2_3_9", + "GroupNorm_10.out13_10.out10_0_9", + "NhwcConv_10_out-/decoder/up_blocks.0/resnets.2/conv2/Conv_output_0.out2_3_10", + "/decoder/up_blocks.0/resnets.2/Add_output_0.nhwc6_0.out_7_1_5", + "/decoder/up_blocks.0/upsamplers.0/Resize_output_0.nhwc6_0.out_6_0", + "NhwcConv_11_out-/decoder/up_blocks.0/upsamplers.0/conv/Conv_output_0.out2_3_11", + "GroupNorm_11.out13_11.out10_0_10", + "NhwcConv_12_out-/decoder/up_blocks.1/resnets.0/conv1/Conv_output_0.out2_3_12", + "GroupNorm_12.out13_12.out10_0_11", + "NhwcConv_13_out-/decoder/up_blocks.1/resnets.0/conv2/Conv_output_0.out2_3_13", + "/decoder/up_blocks.1/resnets.0/Add.out_7_1_6", + "GroupNorm_13.out13_13.out10_0_12", + "NhwcConv_14_out-/decoder/up_blocks.1/resnets.1/conv1/Conv_output_0.out2_3_14", + "GroupNorm_14.out13_14.out10_0_13", + "NhwcConv_15_out-/decoder/up_blocks.1/resnets.1/conv2/Conv_output_0.out2_3_15", + "/decoder/up_blocks.1/resnets.1/Add.out_7_1_7", + "GroupNorm_15.out13_15.out10_0_14", + "NhwcConv_16_out-/decoder/up_blocks.1/resnets.2/conv1/Conv_output_0.out2_3_16", + "GroupNorm_16.out13_16.out10_0_15", + "NhwcConv_17_out-/decoder/up_blocks.1/resnets.2/conv2/Conv_output_0.out2_3_17", + "/decoder/up_blocks.1/resnets.2/Add.out_7_1_8", + "/decoder/up_blocks.1/upsamplers.0/Resize_output_0.nhwc6_1.out_6_1", + "NhwcConv_18_out-/decoder/up_blocks.1/upsamplers.0/conv/Conv_output_0.out2_3_18", + "GroupNorm_17.out13_17.out10_0_16", + "NhwcConv_19_out-/decoder/up_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out2_3_19", + "NhwcConv_20_out-/decoder/up_blocks.2/resnets.0/conv1/Conv_output_0.out2_3_20", + "GroupNorm_18.out13_18.out10_0_17", + "NhwcConv_21_out-/decoder/up_blocks.2/resnets.0/conv2/Conv_output_0.out2_3_21", + "/decoder/up_blocks.2/resnets.0/Add.out_7_1_9", + "GroupNorm_19.out13_19.out10_0_18", + "NhwcConv_22_out-/decoder/up_blocks.2/resnets.1/conv1/Conv_output_0.out2_3_22", + "GroupNorm_20.out13_20.out10_0_19", + "NhwcConv_23_out-/decoder/up_blocks.2/resnets.1/conv2/Conv_output_0.out2_3_23", + "/decoder/up_blocks.2/resnets.1/Add.out_7_1_10", + "GroupNorm_21.out13_21.out10_0_20", + "NhwcConv_24_out-/decoder/up_blocks.2/resnets.2/conv1/Conv_output_0.out2_3_24", + "GroupNorm_22.out13_22.out10_0_21", + "NhwcConv_25_out-/decoder/up_blocks.2/resnets.2/conv2/Conv_output_0.out2_3_25", + "/decoder/up_blocks.2/resnets.2/Add.out_7_1_11", + "/decoder/up_blocks.2/upsamplers.0/Resize_output_0.nhwc6_2.out_6_2", + "NhwcConv_26_out-/decoder/up_blocks.2/upsamplers.0/conv/Conv_output_0.out2_3_26", + "GroupNorm_23.out13_23.out10_0_22", + "NhwcConv_27_out-/decoder/up_blocks.3/resnets.0/conv_shortcut/Conv_output_0.out2_3_27", + "NhwcConv_28_out-/decoder/up_blocks.3/resnets.0/conv1/Conv_output_0.out2_3_28", + "GroupNorm_24.out13_24.out10_0_23", + "NhwcConv_29_out-/decoder/up_blocks.3/resnets.0/conv2/Conv_output_0.out2_3_29", + "/decoder/up_blocks.3/resnets.0/Add.out_7_1_12", + "GroupNorm_25.out13_25.out10_0_24", + "NhwcConv_30_out-/decoder/up_blocks.3/resnets.1/conv1/Conv_output_0.out2_3_30", + "GroupNorm_26.out13_26.out10_0_25", + "NhwcConv_31_out-/decoder/up_blocks.3/resnets.1/conv2/Conv_output_0.out2_3_31", + "/decoder/up_blocks.3/resnets.1/Add.out_7_1_13", + "GroupNorm_27.out13_27.out10_0_26", + "NhwcConv_32_out-/decoder/up_blocks.3/resnets.2/conv1/Conv_output_0.out2_3_32", + "GroupNorm_28.out13_28.out10_0_27", + "NhwcConv_33_out-/decoder/up_blocks.3/resnets.2/conv2/Conv_output_0.out2_3_33", + "/decoder/up_blocks.3/resnets.2/Add.out_7_1_14", + "GroupNorm_29.out13_29.out10_0_28" + ] + }, + "const": { + "buffer_size": 61155328, + "xrt_arg_id": 3, + "packed_tensors": [ + "NhwcConv_0_weight_NHWC", + "GroupNorm_0_wts_10_0_0", + "NhwcConv_1_weight_NHWC", + "GroupNorm_1_wts_10_0_1", + "NhwcConv_2_weight_NHWC", + "GroupNorm_2_wts_10_1_0", + "onnx::MatMul_914", + "onnx::MatMul_913", + "onnx::MatMul_915", + "/decoder/mid_block/attentions.0/MatMulmha_4_0_0_mask.9_0", + "onnx::MatMul_923", + "GroupNorm_3_wts_10_0_2", + "NhwcConv_3_weight_NHWC", + "GroupNorm_4_wts_10_0_3", + "NhwcConv_4_weight_NHWC", + "GroupNorm_5_wts_10_0_4", + "NhwcConv_5_weight_NHWC", + "GroupNorm_6_wts_10_0_5", + "NhwcConv_6_weight_NHWC", + "GroupNorm_7_wts_10_0_6", + "NhwcConv_7_weight_NHWC", + "GroupNorm_8_wts_10_0_7", + "NhwcConv_8_weight_NHWC", + "GroupNorm_9_wts_10_0_8", + "NhwcConv_9_weight_NHWC", + "GroupNorm_10_wts_10_0_9", + "NhwcConv_10_weight_NHWC", + "/decoder/up_blocks.0/upsamplers.0/Resize.weights6_0", + "NhwcConv_11_weight_NHWC", + "GroupNorm_11_wts_10_0_10", + "NhwcConv_12_weight_NHWC", + "GroupNorm_12_wts_10_0_11", + "NhwcConv_13_weight_NHWC", + "GroupNorm_13_wts_10_0_12", + "NhwcConv_14_weight_NHWC", + "GroupNorm_14_wts_10_0_13", + "NhwcConv_15_weight_NHWC", + "GroupNorm_15_wts_10_0_14", + "NhwcConv_16_weight_NHWC", + "GroupNorm_16_wts_10_0_15", + "NhwcConv_17_weight_NHWC", + "/decoder/up_blocks.1/upsamplers.0/Resize.weights6_1", + "NhwcConv_18_weight_NHWC", + "GroupNorm_17_wts_10_0_16", + "NhwcConv_19_weight_NHWC", + "NhwcConv_20_weight_NHWC", + "GroupNorm_18_wts_10_0_17", + "NhwcConv_21_weight_NHWC", + "GroupNorm_19_wts_10_0_18", + "NhwcConv_22_weight_NHWC", + "GroupNorm_20_wts_10_0_19", + "NhwcConv_23_weight_NHWC", + "GroupNorm_21_wts_10_0_20", + "NhwcConv_24_weight_NHWC", + "GroupNorm_22_wts_10_0_21", + "NhwcConv_25_weight_NHWC", + "/decoder/up_blocks.2/upsamplers.0/Resize.weights6_2", + "NhwcConv_26_weight_NHWC", + "GroupNorm_23_wts_10_0_22", + "NhwcConv_27_weight_NHWC", + "NhwcConv_28_weight_NHWC", + "GroupNorm_24_wts_10_0_23", + "NhwcConv_29_weight_NHWC", + "GroupNorm_25_wts_10_0_24", + "NhwcConv_30_weight_NHWC", + "GroupNorm_26_wts_10_0_25", + "NhwcConv_31_weight_NHWC", + "GroupNorm_27_wts_10_0_26", + "NhwcConv_32_weight_NHWC", + "GroupNorm_28_wts_10_0_27", + "NhwcConv_33_weight_NHWC", + "GroupNorm_29_wts_10_0_28", + "NhwcConv_34_weight_NHWC" + ] + }, + "super_instr": { + "buffer_size": 0, + "xrt_arg_id": 4, + "packed_tensors": [] + } + }, + "tensor_map": { + "Transpose_0_out-latent_sample.out2_3_0": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 16 + ], + "size_in_bytes": 32, + "op_tensor_size": 32, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 0 + }, + "NhwcConv_34_out-sample.out2_3_34": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 3 + ], + "size_in_bytes": 6, + "op_tensor_size": 6, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "False" + ], + "offset": 0 + }, + "NhwcConv_0_out-/decoder/conv_in/Conv_output_0.out2_3_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 0 + }, + "GroupNorm_0.out13_0.out10_0_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 1024 + }, + "NhwcConv_1_out-/decoder/mid_block/resnets.0/conv1/Conv_output_0.out2_3_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 2048 + }, + "GroupNorm_1.out13_1.out10_0_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 3072 + }, + "NhwcConv_2_out-/decoder/mid_block/resnets.0/conv2/Conv_output_0.out2_3_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 4096 + }, + "/decoder/mid_block/resnets.0/Add.out_7_1_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 5120 + }, + "/decoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out10_1_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height*width", + "False" + ], + "offset": 6144 + }, + "/decoder/mid_block/attentions.0/to_k/Add_output_0.out3_4_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height*width", + "False" + ], + "offset": 7168 + }, + "/decoder/mid_block/attentions.0/to_q/Add_output_0.out3_4_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height*width", + "False" + ], + "offset": 8192 + }, + "/decoder/mid_block/attentions.0/to_v/Add_output_0.out3_4_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height*width", + "False" + ], + "offset": 9216 + }, + "/decoder/mid_block/attentions.0/Reshape_4_output_0.out9_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height*width", + "False" + ], + "offset": 10240 + }, + "/decoder/mid_block/attentions.0/Transpose_6_output_0_4d.out3_1_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 11264 + }, + "/decoder/mid_block/attentions.0/Add_output_0_NHWC.out_7_1_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 12288 + }, + "GroupNorm_3.out13_3.out10_0_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 13312 + }, + "NhwcConv_3_out-/decoder/mid_block/resnets.1/conv1/Conv_output_0.out2_3_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 14336 + }, + "GroupNorm_4.out13_4.out10_0_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 15360 + }, + "NhwcConv_4_out-/decoder/mid_block/resnets.1/conv2/Conv_output_0.out2_3_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 16384 + }, + "/decoder/mid_block/resnets.1/Add_output_0_NHWC.out_7_1_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 17408 + }, + "GroupNorm_5.out13_5.out10_0_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 18432 + }, + "NhwcConv_5_out-/decoder/up_blocks.0/resnets.0/conv1/Conv_output_0.out2_3_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 19456 + }, + "GroupNorm_6.out13_6.out10_0_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 20480 + }, + "NhwcConv_6_out-/decoder/up_blocks.0/resnets.0/conv2/Conv_output_0.out2_3_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 21504 + }, + "/decoder/up_blocks.0/resnets.0/Add_output_0_NHWC.out_7_1_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 22528 + }, + "GroupNorm_7.out13_7.out10_0_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 23552 + }, + "NhwcConv_7_out-/decoder/up_blocks.0/resnets.1/conv1/Conv_output_0.out2_3_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 24576 + }, + "GroupNorm_8.out13_8.out10_0_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 25600 + }, + "NhwcConv_8_out-/decoder/up_blocks.0/resnets.1/conv2/Conv_output_0.out2_3_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 26624 + }, + "/decoder/up_blocks.0/resnets.1/Add_output_0_NHWC.out_7_1_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 27648 + }, + "GroupNorm_9.out13_9.out10_0_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 28672 + }, + "NhwcConv_9_out-/decoder/up_blocks.0/resnets.2/conv1/Conv_output_0.out2_3_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 29696 + }, + "GroupNorm_10.out13_10.out10_0_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 30720 + }, + "NhwcConv_10_out-/decoder/up_blocks.0/resnets.2/conv2/Conv_output_0.out2_3_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 31744 + }, + "/decoder/up_blocks.0/resnets.2/Add_output_0.nhwc6_0.out_7_1_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "height", + "width", + "False" + ], + "offset": 32768 + }, + "/decoder/up_blocks.0/upsamplers.0/Resize_output_0.nhwc6_0.out_6_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "floor(2.0*height)", + "floor(2.0*width)", + "False" + ], + "offset": 33792 + }, + "NhwcConv_11_out-/decoder/up_blocks.0/upsamplers.0/conv/Conv_output_0.out2_3_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "floor(2.0*height)", + "floor(2.0*width)", + "False" + ], + "offset": 34816 + }, + "GroupNorm_11.out13_11.out10_0_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "floor(2.0*height)", + "floor(2.0*width)", + "False" + ], + "offset": 35840 + }, + "NhwcConv_12_out-/decoder/up_blocks.1/resnets.0/conv1/Conv_output_0.out2_3_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "floor(2.0*height)", + "floor(2.0*width)", + "False" + ], + "offset": 36864 + }, + "GroupNorm_12.out13_12.out10_0_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "floor(2.0*height)", + "floor(2.0*width)", + "False" + ], + "offset": 37888 + }, + "NhwcConv_13_out-/decoder/up_blocks.1/resnets.0/conv2/Conv_output_0.out2_3_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "floor(2.0*height)", + "floor(2.0*width)", + "False" + ], + "offset": 38912 + }, + "/decoder/up_blocks.1/resnets.0/Add.out_7_1_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "floor(2.0*height)", + "floor(2.0*width)", + "False" + ], + "offset": 39936 + }, + "GroupNorm_13.out13_13.out10_0_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "floor(2.0*height)", + "floor(2.0*width)", + "False" + ], + "offset": 40960 + }, + "NhwcConv_14_out-/decoder/up_blocks.1/resnets.1/conv1/Conv_output_0.out2_3_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "floor(2.0*height)", + "floor(2.0*width)", + "False" + ], + "offset": 41984 + }, + "GroupNorm_14.out13_14.out10_0_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "floor(2.0*height)", + "floor(2.0*width)", + "False" + ], + "offset": 43008 + }, + "NhwcConv_15_out-/decoder/up_blocks.1/resnets.1/conv2/Conv_output_0.out2_3_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "floor(2.0*height)", + "floor(2.0*width)", + "False" + ], + "offset": 44032 + }, + "/decoder/up_blocks.1/resnets.1/Add.out_7_1_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "floor(2.0*height)", + "floor(2.0*width)", + "False" + ], + "offset": 45056 + }, + "GroupNorm_15.out13_15.out10_0_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "floor(2.0*height)", + "floor(2.0*width)", + "False" + ], + "offset": 46080 + }, + "NhwcConv_16_out-/decoder/up_blocks.1/resnets.2/conv1/Conv_output_0.out2_3_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "floor(2.0*height)", + "floor(2.0*width)", + "False" + ], + "offset": 47104 + }, + "GroupNorm_16.out13_16.out10_0_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "floor(2.0*height)", + "floor(2.0*width)", + "False" + ], + "offset": 48128 + }, + "NhwcConv_17_out-/decoder/up_blocks.1/resnets.2/conv2/Conv_output_0.out2_3_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "floor(2.0*height)", + "floor(2.0*width)", + "False" + ], + "offset": 49152 + }, + "/decoder/up_blocks.1/resnets.2/Add.out_7_1_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "floor(2.0*height)", + "floor(2.0*width)", + "False" + ], + "offset": 50176 + }, + "/decoder/up_blocks.1/upsamplers.0/Resize_output_0.nhwc6_1.out_6_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "False" + ], + "offset": 51200 + }, + "NhwcConv_18_out-/decoder/up_blocks.1/upsamplers.0/conv/Conv_output_0.out2_3_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "False" + ], + "offset": 52224 + }, + "GroupNorm_17.out13_17.out10_0_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "False" + ], + "offset": 53248 + }, + "NhwcConv_19_out-/decoder/up_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out2_3_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "False" + ], + "offset": 54272 + }, + "NhwcConv_20_out-/decoder/up_blocks.2/resnets.0/conv1/Conv_output_0.out2_3_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "False" + ], + "offset": 54784 + }, + "GroupNorm_18.out13_18.out10_0_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "False" + ], + "offset": 55296 + }, + "NhwcConv_21_out-/decoder/up_blocks.2/resnets.0/conv2/Conv_output_0.out2_3_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "False" + ], + "offset": 55808 + }, + "/decoder/up_blocks.2/resnets.0/Add.out_7_1_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "False" + ], + "offset": 56320 + }, + "GroupNorm_19.out13_19.out10_0_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "False" + ], + "offset": 56832 + }, + "NhwcConv_22_out-/decoder/up_blocks.2/resnets.1/conv1/Conv_output_0.out2_3_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "False" + ], + "offset": 57344 + }, + "GroupNorm_20.out13_20.out10_0_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "False" + ], + "offset": 57856 + }, + "NhwcConv_23_out-/decoder/up_blocks.2/resnets.1/conv2/Conv_output_0.out2_3_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "False" + ], + "offset": 58368 + }, + "/decoder/up_blocks.2/resnets.1/Add.out_7_1_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "False" + ], + "offset": 58880 + }, + "GroupNorm_21.out13_21.out10_0_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "False" + ], + "offset": 59392 + }, + "NhwcConv_24_out-/decoder/up_blocks.2/resnets.2/conv1/Conv_output_0.out2_3_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "False" + ], + "offset": 59904 + }, + "GroupNorm_22.out13_22.out10_0_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "False" + ], + "offset": 60416 + }, + "NhwcConv_25_out-/decoder/up_blocks.2/resnets.2/conv2/Conv_output_0.out2_3_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "False" + ], + "offset": 60928 + }, + "/decoder/up_blocks.2/resnets.2/Add.out_7_1_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*height))", + "floor(2.0*floor(2.0*width))", + "False" + ], + "offset": 61440 + }, + "/decoder/up_blocks.2/upsamplers.0/Resize_output_0.nhwc6_2.out_6_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "False" + ], + "offset": 61952 + }, + "NhwcConv_26_out-/decoder/up_blocks.2/upsamplers.0/conv/Conv_output_0.out2_3_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "False" + ], + "offset": 62464 + }, + "GroupNorm_23.out13_23.out10_0_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "False" + ], + "offset": 62976 + }, + "NhwcConv_27_out-/decoder/up_blocks.3/resnets.0/conv_shortcut/Conv_output_0.out2_3_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "False" + ], + "offset": 63488 + }, + "NhwcConv_28_out-/decoder/up_blocks.3/resnets.0/conv1/Conv_output_0.out2_3_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "False" + ], + "offset": 63744 + }, + "GroupNorm_24.out13_24.out10_0_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "False" + ], + "offset": 64000 + }, + "NhwcConv_29_out-/decoder/up_blocks.3/resnets.0/conv2/Conv_output_0.out2_3_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "False" + ], + "offset": 64256 + }, + "/decoder/up_blocks.3/resnets.0/Add.out_7_1_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "False" + ], + "offset": 64512 + }, + "GroupNorm_25.out13_25.out10_0_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "False" + ], + "offset": 64768 + }, + "NhwcConv_30_out-/decoder/up_blocks.3/resnets.1/conv1/Conv_output_0.out2_3_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "False" + ], + "offset": 65024 + }, + "GroupNorm_26.out13_26.out10_0_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "False" + ], + "offset": 65280 + }, + "NhwcConv_31_out-/decoder/up_blocks.3/resnets.1/conv2/Conv_output_0.out2_3_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "False" + ], + "offset": 65536 + }, + "/decoder/up_blocks.3/resnets.1/Add.out_7_1_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "False" + ], + "offset": 65792 + }, + "GroupNorm_27.out13_27.out10_0_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "False" + ], + "offset": 66048 + }, + "NhwcConv_32_out-/decoder/up_blocks.3/resnets.2/conv1/Conv_output_0.out2_3_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "False" + ], + "offset": 66304 + }, + "GroupNorm_28.out13_28.out10_0_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "False" + ], + "offset": 66560 + }, + "NhwcConv_33_out-/decoder/up_blocks.3/resnets.2/conv2/Conv_output_0.out2_3_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "False" + ], + "offset": 66816 + }, + "/decoder/up_blocks.3/resnets.2/Add.out_7_1_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "False" + ], + "offset": 67072 + }, + "GroupNorm_29.out13_29.out10_0_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 1, + 1, + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "dynamic_shapes": [ + "False", + "floor(2.0*floor(2.0*floor(2.0*height)))", + "floor(2.0*floor(2.0*floor(2.0*width)))", + "False" + ], + "offset": 67328 + }, + "NhwcConv_0_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 91136 + ], + "size_in_bytes": 91136, + "op_tensor_size": 91136, + "offset": 0, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_0.const", + "file_size": 91136 + }, + "GroupNorm_0_wts_10_0_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 91136, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_1.const", + "file_size": 2048 + }, + "NhwcConv_1_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 93184, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_2.const", + "file_size": 2916352 + }, + "GroupNorm_1_wts_10_0_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 3009536, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_3.const", + "file_size": 2048 + }, + "NhwcConv_2_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 3011584, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_4.const", + "file_size": 2916352 + }, + "GroupNorm_2_wts_10_1_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 5927936, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_5.const", + "file_size": 2048 + }, + "onnx::MatMul_914": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 299008 + ], + "size_in_bytes": 299008, + "op_tensor_size": 299008, + "offset": 5929984, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_6.const", + "file_size": 299008 + }, + "onnx::MatMul_913": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 299008 + ], + "size_in_bytes": 299008, + "op_tensor_size": 299008, + "offset": 6228992, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_7.const", + "file_size": 299008 + }, + "onnx::MatMul_915": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 299008 + ], + "size_in_bytes": 299008, + "op_tensor_size": 299008, + "offset": 6528000, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_8.const", + "file_size": 299008 + }, + "/decoder/mid_block/attentions.0/MatMulmha_4_0_0_mask.9_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 16384 + ], + "size_in_bytes": 32768, + "op_tensor_size": 32768, + "offset": 6827008, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_9.const", + "file_size": 32768 + }, + "onnx::MatMul_923": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 299008 + ], + "size_in_bytes": 299008, + "op_tensor_size": 299008, + "offset": 6859776, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_10.const", + "file_size": 299008 + }, + "GroupNorm_3_wts_10_0_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 7158784, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_11.const", + "file_size": 2048 + }, + "NhwcConv_3_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 7160832, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_12.const", + "file_size": 2916352 + }, + "GroupNorm_4_wts_10_0_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 10077184, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_13.const", + "file_size": 2048 + }, + "NhwcConv_4_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 10079232, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_14.const", + "file_size": 2916352 + }, + "GroupNorm_5_wts_10_0_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 12995584, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_15.const", + "file_size": 2048 + }, + "NhwcConv_5_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 12997632, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_16.const", + "file_size": 2916352 + }, + "GroupNorm_6_wts_10_0_5": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 15913984, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_17.const", + "file_size": 2048 + }, + "NhwcConv_6_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 15916032, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_18.const", + "file_size": 2916352 + }, + "GroupNorm_7_wts_10_0_6": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 18832384, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_19.const", + "file_size": 2048 + }, + "NhwcConv_7_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 18834432, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_20.const", + "file_size": 2916352 + }, + "GroupNorm_8_wts_10_0_7": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 21750784, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_21.const", + "file_size": 2048 + }, + "NhwcConv_8_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 21752832, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_22.const", + "file_size": 2916352 + }, + "GroupNorm_9_wts_10_0_8": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 24669184, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_23.const", + "file_size": 2048 + }, + "NhwcConv_9_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 24671232, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_24.const", + "file_size": 2916352 + }, + "GroupNorm_10_wts_10_0_9": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 27587584, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_25.const", + "file_size": 2048 + }, + "NhwcConv_10_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 27589632, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_26.const", + "file_size": 2916352 + }, + "/decoder/up_blocks.0/upsamplers.0/Resize.weights6_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "offset": 30505984, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_27.const", + "file_size": 256 + }, + "NhwcConv_11_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 30506240, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_28.const", + "file_size": 2916352 + }, + "GroupNorm_11_wts_10_0_10": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 33422592, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_29.const", + "file_size": 2048 + }, + "NhwcConv_12_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 33424640, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_30.const", + "file_size": 2916352 + }, + "GroupNorm_12_wts_10_0_11": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 36340992, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_31.const", + "file_size": 2048 + }, + "NhwcConv_13_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 36343040, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_32.const", + "file_size": 2916352 + }, + "GroupNorm_13_wts_10_0_12": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 39259392, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_33.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": 39261440, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_34.const", + "file_size": 2916352 + }, + "GroupNorm_14_wts_10_0_13": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 42177792, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_35.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": 42179840, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_36.const", + "file_size": 2916352 + }, + "GroupNorm_15_wts_10_0_14": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 45096192, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_37.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": 45098240, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_38.const", + "file_size": 2916352 + }, + "GroupNorm_16_wts_10_0_15": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 48014592, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_39.const", + "file_size": 2048 + }, + "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": 48016640, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_40.const", + "file_size": 2916352 + }, + "/decoder/up_blocks.1/upsamplers.0/Resize.weights6_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "offset": 50932992, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_41.const", + "file_size": 256 + }, + "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": 50933248, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_42.const", + "file_size": 2916352 + }, + "GroupNorm_17_wts_10_0_16": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 53849600, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_43.const", + "file_size": 2048 + }, + "NhwcConv_19_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 163840 + ], + "size_in_bytes": 163840, + "op_tensor_size": 163840, + "offset": 53851648, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_44.const", + "file_size": 163840 + }, + "NhwcConv_20_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1458176 + ], + "size_in_bytes": 1458176, + "op_tensor_size": 1458176, + "offset": 54015488, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_45.const", + "file_size": 1458176 + }, + "GroupNorm_18_wts_10_0_17": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "offset": 55473664, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_46.const", + "file_size": 1024 + }, + "NhwcConv_21_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 729088 + ], + "size_in_bytes": 729088, + "op_tensor_size": 729088, + "offset": 55474688, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_47.const", + "file_size": 729088 + }, + "GroupNorm_19_wts_10_0_18": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "offset": 56203776, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_48.const", + "file_size": 1024 + }, + "NhwcConv_22_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 729088 + ], + "size_in_bytes": 729088, + "op_tensor_size": 729088, + "offset": 56204800, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_49.const", + "file_size": 729088 + }, + "GroupNorm_20_wts_10_0_19": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "offset": 56933888, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_50.const", + "file_size": 1024 + }, + "NhwcConv_23_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 729088 + ], + "size_in_bytes": 729088, + "op_tensor_size": 729088, + "offset": 56934912, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_51.const", + "file_size": 729088 + }, + "GroupNorm_21_wts_10_0_20": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "offset": 57664000, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_52.const", + "file_size": 1024 + }, + "NhwcConv_24_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 729088 + ], + "size_in_bytes": 729088, + "op_tensor_size": 729088, + "offset": 57665024, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_53.const", + "file_size": 729088 + }, + "GroupNorm_22_wts_10_0_21": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "offset": 58394112, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_54.const", + "file_size": 1024 + }, + "NhwcConv_25_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 729088 + ], + "size_in_bytes": 729088, + "op_tensor_size": 729088, + "offset": 58395136, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_55.const", + "file_size": 729088 + }, + "/decoder/up_blocks.2/upsamplers.0/Resize.weights6_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "offset": 59124224, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_56.const", + "file_size": 256 + }, + "NhwcConv_26_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 729088 + ], + "size_in_bytes": 729088, + "op_tensor_size": 729088, + "offset": 59124480, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_57.const", + "file_size": 729088 + }, + "GroupNorm_23_wts_10_0_22": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "offset": 59853568, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_58.const", + "file_size": 1024 + }, + "NhwcConv_27_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 40960 + ], + "size_in_bytes": 40960, + "op_tensor_size": 40960, + "offset": 59854592, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_59.const", + "file_size": 40960 + }, + "NhwcConv_28_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 364544 + ], + "size_in_bytes": 364544, + "op_tensor_size": 364544, + "offset": 59895552, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_60.const", + "file_size": 364544 + }, + "GroupNorm_24_wts_10_0_23": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "offset": 60260096, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_61.const", + "file_size": 512 + }, + "NhwcConv_29_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 174080 + ], + "size_in_bytes": 174080, + "op_tensor_size": 174080, + "offset": 60260608, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_62.const", + "file_size": 174080 + }, + "GroupNorm_25_wts_10_0_24": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "offset": 60434688, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_63.const", + "file_size": 512 + }, + "NhwcConv_30_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 174080 + ], + "size_in_bytes": 174080, + "op_tensor_size": 174080, + "offset": 60435200, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_64.const", + "file_size": 174080 + }, + "GroupNorm_26_wts_10_0_25": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "offset": 60609280, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_65.const", + "file_size": 512 + }, + "NhwcConv_31_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 174080 + ], + "size_in_bytes": 174080, + "op_tensor_size": 174080, + "offset": 60609792, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_66.const", + "file_size": 174080 + }, + "GroupNorm_27_wts_10_0_26": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "offset": 60783872, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_67.const", + "file_size": 512 + }, + "NhwcConv_32_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 174080 + ], + "size_in_bytes": 174080, + "op_tensor_size": 174080, + "offset": 60784384, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_68.const", + "file_size": 174080 + }, + "GroupNorm_28_wts_10_0_27": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "offset": 60958464, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_69.const", + "file_size": 512 + }, + "NhwcConv_33_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 174080 + ], + "size_in_bytes": 174080, + "op_tensor_size": 174080, + "offset": 60958976, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_70.const", + "file_size": 174080 + }, + "GroupNorm_29_wts_10_0_28": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "offset": 61133056, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_71.const", + "file_size": 512 + }, + "NhwcConv_34_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 21760 + ], + "size_in_bytes": 21760, + "op_tensor_size": 21760, + "offset": 61133568, + "file_name": "cache/NhwcConv_0-decoderconv_inConv_72.const", + "file_size": 21760 + } + }, + "dynamic_shape_subgraph": true, + "dynamic_shape_list": [ + { + "floor(2.0*floor(2.0*floor(2.0*width)))": 512, + "width": 64, + "floor(2.0*width)": 128, + "floor(2.0*height)": 128, + "floor(2.0*floor(2.0*floor(2.0*height)))": 512, + "floor(2.0*floor(2.0*height))": 256, + "height*width": 4096, + "height": 64, + "floor(2.0*floor(2.0*width))": 256 + }, + { + "floor(2.0*floor(2.0*floor(2.0*width)))": 768, + "width": 96, + "floor(2.0*width)": 192, + "floor(2.0*height)": 128, + "floor(2.0*floor(2.0*floor(2.0*height)))": 512, + "floor(2.0*floor(2.0*height))": 256, + "height*width": 6144, + "height": 64, + "floor(2.0*floor(2.0*width))": 384 + }, + { + "floor(2.0*floor(2.0*floor(2.0*width)))": 512, + "width": 64, + "floor(2.0*width)": 128, + "floor(2.0*height)": 192, + "floor(2.0*floor(2.0*floor(2.0*height)))": 768, + "floor(2.0*floor(2.0*height))": 384, + "height*width": 6144, + "height": 96, + "floor(2.0*floor(2.0*width))": 256 + }, + { + "floor(2.0*floor(2.0*floor(2.0*width)))": 1024, + "width": 128, + "floor(2.0*width)": 256, + "floor(2.0*height)": 144, + "floor(2.0*floor(2.0*floor(2.0*height)))": 576, + "floor(2.0*floor(2.0*height))": 288, + "height*width": 9216, + "height": 72, + "floor(2.0*floor(2.0*width))": 512 + }, + { + "floor(2.0*floor(2.0*floor(2.0*width)))": 576, + "width": 72, + "floor(2.0*width)": 144, + "floor(2.0*height)": 256, + "floor(2.0*floor(2.0*floor(2.0*height)))": 1024, + "floor(2.0*floor(2.0*height))": 512, + "height*width": 9216, + "height": 128, + "floor(2.0*floor(2.0*width))": 288 + }, + { + "floor(2.0*floor(2.0*floor(2.0*width)))": 1024, + "width": 128, + "floor(2.0*width)": 256, + "floor(2.0*height)": 192, + "floor(2.0*floor(2.0*floor(2.0*height)))": 768, + "floor(2.0*floor(2.0*height))": 384, + "height*width": 12288, + "height": 96, + "floor(2.0*floor(2.0*width))": 512 + }, + { + "floor(2.0*floor(2.0*floor(2.0*width)))": 768, + "width": 96, + "floor(2.0*width)": 192, + "floor(2.0*height)": 256, + "floor(2.0*floor(2.0*floor(2.0*height)))": 1024, + "floor(2.0*floor(2.0*height))": 512, + "height*width": 12288, + "height": 128, + "floor(2.0*floor(2.0*width))": 384 + }, + { + "floor(2.0*floor(2.0*floor(2.0*width)))": 1024, + "width": 128, + "floor(2.0*width)": 256, + "floor(2.0*height)": 256, + "floor(2.0*floor(2.0*floor(2.0*height)))": 1024, + "floor(2.0*floor(2.0*height))": 512, + "height*width": 16384, + "height": 128, + "floor(2.0*floor(2.0*width))": 512 + } + ], + "aux_info": {} +} \ No newline at end of file