diff --git "a/vae_decoder/dd/cache/NhwcConv_0-post_quant_convConv_meta.json" "b/vae_decoder/dd/cache/NhwcConv_0-post_quant_convConv_meta.json" new file mode 100644--- /dev/null +++ "b/vae_decoder/dd/cache/NhwcConv_0-post_quant_convConv_meta.json" @@ -0,0 +1,10622 @@ +{ + "dd_meta_major_version": 1, + "dd_meta_minor_version": 4, + "state_table_updates": [], + "op_list": [ + { + "name": "NhwcConv_0-/post_quant_conv/Conv", + "type": "SDConv", + "in_args": [ + "Transpose_0_out-latents.out0_1_0" + ], + "const_args": [ + "NhwcConv_0_weight_NHWC" + ], + "out_args": [ + "NhwcConv_0_out-/post_quant_conv/Conv_output_0.out0_1_0" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "4" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "4" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "4", + "1", + "1", + "4" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "NhwcConv_1-/decoder/conv_in/Conv", + "type": "SDConv", + "in_args": [ + "NhwcConv_0_out-/post_quant_conv/Conv_output_0.out0_1_0" + ], + "const_args": [ + "NhwcConv_1_weight_NHWC" + ], + "out_args": [ + "NhwcConv_1_out-/decoder/conv_in/Conv_output_0.out0_1_1" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "4" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "4" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_0", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_1_out-/decoder/conv_in/Conv_output_0.out0_1_1" + ], + "const_args": [ + "GroupNorm_0_wts_4_0_0" + ], + "out_args": [ + "GroupNorm_0.out12_0.out4_0_0" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "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/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_0.out12_0.out4_0_0" + ], + "const_args": [ + "NhwcConv_2_weight_NHWC" + ], + "out_args": [ + "NhwcConv_2_out-/decoder/mid_block/resnets.0/conv1/Conv_output_0.out0_1_2" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_1", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_2_out-/decoder/mid_block/resnets.0/conv1/Conv_output_0.out0_1_2" + ], + "const_args": [ + "GroupNorm_1_wts_4_0_1" + ], + "out_args": [ + "GroupNorm_1.out12_1.out4_0_1" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "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.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_1.out12_1.out4_0_1" + ], + "const_args": [ + "NhwcConv_3_weight_NHWC" + ], + "out_args": [ + "NhwcConv_3_out-/decoder/mid_block/resnets.0/conv2/Conv_output_0.out0_1_3" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/decoder/mid_block/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_1_out-/decoder/conv_in/Conv_output_0.out0_1_1", + "NhwcConv_3_out-/decoder/mid_block/resnets.0/conv2/Conv_output_0.out0_1_3" + ], + "const_args": [], + "out_args": [ + "/decoder/mid_block/resnets.0/Add.out_2_1_0" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "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_2_1_0" + ], + "const_args": [ + "GroupNorm_2_wts_4_1_0" + ], + "out_args": [ + "/decoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out4_1_0" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "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/MatMulAddFusion", + "type": "SDGemm", + "in_args": [ + "/decoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out4_1_0" + ], + "const_args": [ + "onnx::MatMul_918_2_0_0" + ], + "out_args": [ + "/decoder/mid_block/attentions.0/Reshape_2_output_0_3d.out1_3_0" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "1", + "4096", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "4096", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + } + } + }, + { + "name": "/decoder/mid_block/attentions.0/to_q/MatMul/MatMulAddFusion", + "type": "SDGemm", + "in_args": [ + "/decoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out4_1_0" + ], + "const_args": [ + "onnx::MatMul_917_2_0_1" + ], + "out_args": [ + "/decoder/mid_block/attentions.0/Reshape_1_output_0_3d.out1_3_1" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "1", + "4096", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "4096", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + } + } + }, + { + "name": "/decoder/mid_block/attentions.0/to_v/MatMul/MatMulAddFusion", + "type": "SDGemm", + "in_args": [ + "/decoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out4_1_0" + ], + "const_args": [ + "onnx::MatMul_919_2_0_2" + ], + "out_args": [ + "/decoder/mid_block/attentions.0/Reshape_3_output_0_3d.out1_3_2" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "1", + "4096", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "4096", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + } + } + }, + { + "name": "/decoder/mid_block/attentions.0/MatMul/MatMulScaleFusion/_FusedActivationmha_3_1_0", + "type": "SDMHA_VAE", + "in_args": [ + "/decoder/mid_block/attentions.0/Reshape_1_output_0_3d.out1_3_1", + "/decoder/mid_block/attentions.0/Reshape_2_output_0_3d.out1_3_0", + "/decoder/mid_block/attentions.0/Reshape_3_output_0_3d.out1_3_2" + ], + "const_args": [ + "/decoder/mid_block/attentions.0/MatMul/MatMulScaleFusion/_FusedActivationmha_3_1_0_mask.10_0" + ], + "out_args": [ + "/decoder/mid_block/attentions.0/Transpose_5_output_0_3d.out10_0" + ], + "attrs": { + "num_heads": { + "type": "int", + "value": [ + "1" + ] + }, + "unidirectional": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "4096", + "512", + "4096" + ] + }, + "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/MatMulAddFusion", + "type": "SDGemm", + "in_args": [ + "/decoder/mid_block/attentions.0/Transpose_5_output_0_3d.out10_0" + ], + "const_args": [ + "onnx::MatMul_927_2_0_3" + ], + "out_args": [ + "/decoder/mid_block/attentions.0/Reshape_5_output_0.nhwc6_0.out1_0_0" + ], + "attrs": { + "input_shape": { + "type": "int", + "value": [ + "1", + "4096", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "4096", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "bias_enable": { + "type": "int", + "value": [ + "1" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "bfloat16" + ] + } + } + }, + { + "name": "/decoder/mid_block/attentions.0/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/mid_block/attentions.0/Reshape_5_output_0.nhwc6_0.out1_0_0", + "/decoder/mid_block/resnets.0/Add.out_2_1_0" + ], + "const_args": [], + "out_args": [ + "/decoder/mid_block/attentions.0/Add_output_0_NHWC.out_2_1_1" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "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_2_1_1" + ], + "const_args": [ + "GroupNorm_3_wts_4_0_2" + ], + "out_args": [ + "GroupNorm_3.out12_3.out4_0_2" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "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/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_3.out12_3.out4_0_2" + ], + "const_args": [ + "NhwcConv_4_weight_NHWC" + ], + "out_args": [ + "NhwcConv_4_out-/decoder/mid_block/resnets.1/conv1/Conv_output_0.out0_1_4" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_4", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_4_out-/decoder/mid_block/resnets.1/conv1/Conv_output_0.out0_1_4" + ], + "const_args": [ + "GroupNorm_4_wts_4_0_3" + ], + "out_args": [ + "GroupNorm_4.out12_4.out4_0_3" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "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/mid_block/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_4.out12_4.out4_0_3" + ], + "const_args": [ + "NhwcConv_5_weight_NHWC" + ], + "out_args": [ + "NhwcConv_5_out-/decoder/mid_block/resnets.1/conv2/Conv_output_0.out0_1_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": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/decoder/mid_block/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/mid_block/attentions.0/Add_output_0_NHWC.out_2_1_1", + "NhwcConv_5_out-/decoder/mid_block/resnets.1/conv2/Conv_output_0.out0_1_5" + ], + "const_args": [], + "out_args": [ + "/decoder/mid_block/resnets.1/Add_output_0_NHWC.out_2_1_2" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "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_2_1_2" + ], + "const_args": [ + "GroupNorm_5_wts_4_0_4" + ], + "out_args": [ + "GroupNorm_5.out12_5.out4_0_4" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "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/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_5.out12_5.out4_0_4" + ], + "const_args": [ + "NhwcConv_6_weight_NHWC" + ], + "out_args": [ + "NhwcConv_6_out-/decoder/up_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_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": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_6", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_6_out-/decoder/up_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_6" + ], + "const_args": [ + "GroupNorm_6_wts_4_0_5" + ], + "out_args": [ + "GroupNorm_6.out12_6.out4_0_5" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "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.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_6.out12_6.out4_0_5" + ], + "const_args": [ + "NhwcConv_7_weight_NHWC" + ], + "out_args": [ + "NhwcConv_7_out-/decoder/up_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_7" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/decoder/up_blocks.0/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/mid_block/resnets.1/Add_output_0_NHWC.out_2_1_2", + "NhwcConv_7_out-/decoder/up_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_7" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.0/resnets.0/Add_output_0_NHWC.out_2_1_3" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "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_2_1_3" + ], + "const_args": [ + "GroupNorm_7_wts_4_0_6" + ], + "out_args": [ + "GroupNorm_7.out12_7.out4_0_6" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "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/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_7.out12_7.out4_0_6" + ], + "const_args": [ + "NhwcConv_8_weight_NHWC" + ], + "out_args": [ + "NhwcConv_8_out-/decoder/up_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_8" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_8", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_8_out-/decoder/up_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_8" + ], + "const_args": [ + "GroupNorm_8_wts_4_0_7" + ], + "out_args": [ + "GroupNorm_8.out12_8.out4_0_7" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "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.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_8.out12_8.out4_0_7" + ], + "const_args": [ + "NhwcConv_9_weight_NHWC" + ], + "out_args": [ + "NhwcConv_9_out-/decoder/up_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_9" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/decoder/up_blocks.0/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/up_blocks.0/resnets.0/Add_output_0_NHWC.out_2_1_3", + "NhwcConv_9_out-/decoder/up_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_9" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.0/resnets.1/Add_output_0_NHWC.out_2_1_4" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "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_2_1_4" + ], + "const_args": [ + "GroupNorm_9_wts_4_0_8" + ], + "out_args": [ + "GroupNorm_9.out12_9.out4_0_8" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "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/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_9.out12_9.out4_0_8" + ], + "const_args": [ + "NhwcConv_10_weight_NHWC" + ], + "out_args": [ + "NhwcConv_10_out-/decoder/up_blocks.0/resnets.2/conv1/Conv_output_0.out0_1_10" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_10", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_10_out-/decoder/up_blocks.0/resnets.2/conv1/Conv_output_0.out0_1_10" + ], + "const_args": [ + "GroupNorm_10_wts_4_0_9" + ], + "out_args": [ + "GroupNorm_10.out12_10.out4_0_9" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "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_11-/decoder/up_blocks.0/resnets.2/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_10.out12_10.out4_0_9" + ], + "const_args": [ + "NhwcConv_11_weight_NHWC" + ], + "out_args": [ + "NhwcConv_11_out-/decoder/up_blocks.0/resnets.2/conv2/Conv_output_0.out0_1_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": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/decoder/up_blocks.0/resnets.2/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/up_blocks.0/resnets.1/Add_output_0_NHWC.out_2_1_4", + "NhwcConv_11_out-/decoder/up_blocks.0/resnets.2/conv2/Conv_output_0.out0_1_11" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.0/resnets.2/Add_output_0.nhwc5_0.out_2_1_5" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "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.nhwc5_0.out_2_1_5" + ], + "const_args": [ + "/decoder/up_blocks.0/upsamplers.0/Resize.weights7_0" + ], + "out_args": [ + "/decoder/up_blocks.0/upsamplers.0/Resize_output_0.nhwc5_0.out_7_0" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "64", + "64", + "512" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_12-/decoder/up_blocks.0/upsamplers.0/conv/Conv", + "type": "SDConv", + "in_args": [ + "/decoder/up_blocks.0/upsamplers.0/Resize_output_0.nhwc5_0.out_7_0" + ], + "const_args": [ + "NhwcConv_12_weight_NHWC" + ], + "out_args": [ + "NhwcConv_12_out-/decoder/up_blocks.0/upsamplers.0/conv/Conv_output_0.out0_1_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": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_11", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_12_out-/decoder/up_blocks.0/upsamplers.0/conv/Conv_output_0.out0_1_12" + ], + "const_args": [ + "GroupNorm_11_wts_4_0_10" + ], + "out_args": [ + "GroupNorm_11.out12_11.out4_0_10" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_13-/decoder/up_blocks.1/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_11.out12_11.out4_0_10" + ], + "const_args": [ + "NhwcConv_13_weight_NHWC" + ], + "out_args": [ + "NhwcConv_13_out-/decoder/up_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_13" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_12", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_13_out-/decoder/up_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_13" + ], + "const_args": [ + "GroupNorm_12_wts_4_0_11" + ], + "out_args": [ + "GroupNorm_12.out12_12.out4_0_11" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_14-/decoder/up_blocks.1/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_12.out12_12.out4_0_11" + ], + "const_args": [ + "NhwcConv_14_weight_NHWC" + ], + "out_args": [ + "NhwcConv_14_out-/decoder/up_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_14" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/decoder/up_blocks.1/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_12_out-/decoder/up_blocks.0/upsamplers.0/conv/Conv_output_0.out0_1_12", + "NhwcConv_14_out-/decoder/up_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_14" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.1/resnets.0/Add.out_2_1_6" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_13", + "type": "SDGroupNorm", + "in_args": [ + "/decoder/up_blocks.1/resnets.0/Add.out_2_1_6" + ], + "const_args": [ + "GroupNorm_13_wts_4_0_12" + ], + "out_args": [ + "GroupNorm_13.out12_13.out4_0_12" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_15-/decoder/up_blocks.1/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_13.out12_13.out4_0_12" + ], + "const_args": [ + "NhwcConv_15_weight_NHWC" + ], + "out_args": [ + "NhwcConv_15_out-/decoder/up_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_15" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_14", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_15_out-/decoder/up_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_15" + ], + "const_args": [ + "GroupNorm_14_wts_4_0_13" + ], + "out_args": [ + "GroupNorm_14.out12_14.out4_0_13" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_16-/decoder/up_blocks.1/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_14.out12_14.out4_0_13" + ], + "const_args": [ + "NhwcConv_16_weight_NHWC" + ], + "out_args": [ + "NhwcConv_16_out-/decoder/up_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_16" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/decoder/up_blocks.1/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/up_blocks.1/resnets.0/Add.out_2_1_6", + "NhwcConv_16_out-/decoder/up_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_16" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.1/resnets.1/Add.out_2_1_7" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "GroupNorm_15", + "type": "SDGroupNorm", + "in_args": [ + "/decoder/up_blocks.1/resnets.1/Add.out_2_1_7" + ], + "const_args": [ + "GroupNorm_15_wts_4_0_14" + ], + "out_args": [ + "GroupNorm_15.out12_15.out4_0_14" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_17-/decoder/up_blocks.1/resnets.2/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_15.out12_15.out4_0_14" + ], + "const_args": [ + "NhwcConv_17_weight_NHWC" + ], + "out_args": [ + "NhwcConv_17_out-/decoder/up_blocks.1/resnets.2/conv1/Conv_output_0.out0_1_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": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_16", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_17_out-/decoder/up_blocks.1/resnets.2/conv1/Conv_output_0.out0_1_17" + ], + "const_args": [ + "GroupNorm_16_wts_4_0_15" + ], + "out_args": [ + "GroupNorm_16.out12_16.out4_0_15" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_18-/decoder/up_blocks.1/resnets.2/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_16.out12_16.out4_0_15" + ], + "const_args": [ + "NhwcConv_18_weight_NHWC" + ], + "out_args": [ + "NhwcConv_18_out-/decoder/up_blocks.1/resnets.2/conv2/Conv_output_0.out0_1_18" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/decoder/up_blocks.1/resnets.2/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/up_blocks.1/resnets.1/Add.out_2_1_7", + "NhwcConv_18_out-/decoder/up_blocks.1/resnets.2/conv2/Conv_output_0.out0_1_18" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.1/resnets.2/Add.out_2_1_8" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "is_bias_add": { + "type": "int", + "value": [ + "0" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "/decoder/up_blocks.1/upsamplers.0/Resize", + "type": "SDResize", + "in_args": [ + "/decoder/up_blocks.1/resnets.2/Add.out_2_1_8" + ], + "const_args": [ + "/decoder/up_blocks.1/upsamplers.0/Resize.weights7_1" + ], + "out_args": [ + "/decoder/up_blocks.1/upsamplers.0/Resize_output_0.nhwc5_1.out_7_1" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "128", + "128", + "512" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_19-/decoder/up_blocks.1/upsamplers.0/conv/Conv", + "type": "SDConv", + "in_args": [ + "/decoder/up_blocks.1/upsamplers.0/Resize_output_0.nhwc5_1.out_7_1" + ], + "const_args": [ + "NhwcConv_19_weight_NHWC" + ], + "out_args": [ + "NhwcConv_19_out-/decoder/up_blocks.1/upsamplers.0/conv/Conv_output_0.out0_1_19" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "512", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_17", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_19_out-/decoder/up_blocks.1/upsamplers.0/conv/Conv_output_0.out0_1_19" + ], + "const_args": [ + "GroupNorm_17_wts_4_0_16" + ], + "out_args": [ + "GroupNorm_17.out12_17.out4_0_16" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "1024" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_20-/decoder/up_blocks.2/resnets.0/conv_shortcut/Conv", + "type": "SDConv", + "in_args": [ + "NhwcConv_19_out-/decoder/up_blocks.1/upsamplers.0/conv/Conv_output_0.out0_1_19" + ], + "const_args": [ + "NhwcConv_20_weight_NHWC" + ], + "out_args": [ + "NhwcConv_20_out-/decoder/up_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_20" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "1", + "1", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "NhwcConv_21-/decoder/up_blocks.2/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_17.out12_17.out4_0_16" + ], + "const_args": [ + "NhwcConv_21_weight_NHWC" + ], + "out_args": [ + "NhwcConv_21_out-/decoder/up_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_21" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "512" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "3", + "3", + "512" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_18", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_21_out-/decoder/up_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_21" + ], + "const_args": [ + "GroupNorm_18_wts_4_0_17" + ], + "out_args": [ + "GroupNorm_18.out12_18.out4_0_17" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_22-/decoder/up_blocks.2/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_18.out12_18.out4_0_17" + ], + "const_args": [ + "NhwcConv_22_weight_NHWC" + ], + "out_args": [ + "NhwcConv_22_out-/decoder/up_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_22" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "3", + "3", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/decoder/up_blocks.2/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_20_out-/decoder/up_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_20", + "NhwcConv_22_out-/decoder/up_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_22" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.2/resnets.0/Add.out_2_1_9" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "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_2_1_9" + ], + "const_args": [ + "GroupNorm_19_wts_4_0_18" + ], + "out_args": [ + "GroupNorm_19.out12_19.out4_0_18" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_23-/decoder/up_blocks.2/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_19.out12_19.out4_0_18" + ], + "const_args": [ + "NhwcConv_23_weight_NHWC" + ], + "out_args": [ + "NhwcConv_23_out-/decoder/up_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_23" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "3", + "3", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_20", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_23_out-/decoder/up_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_23" + ], + "const_args": [ + "GroupNorm_20_wts_4_0_19" + ], + "out_args": [ + "GroupNorm_20.out12_20.out4_0_19" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_24-/decoder/up_blocks.2/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_20.out12_20.out4_0_19" + ], + "const_args": [ + "NhwcConv_24_weight_NHWC" + ], + "out_args": [ + "NhwcConv_24_out-/decoder/up_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_24" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "3", + "3", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/decoder/up_blocks.2/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/up_blocks.2/resnets.0/Add.out_2_1_9", + "NhwcConv_24_out-/decoder/up_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_24" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.2/resnets.1/Add.out_2_1_10" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "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_2_1_10" + ], + "const_args": [ + "GroupNorm_21_wts_4_0_20" + ], + "out_args": [ + "GroupNorm_21.out12_21.out4_0_20" + ], + "attrs": { + "epsilon": { + "type": "float", + "value": [ + "9.999999974752427e-07" + ] + }, + "groups": { + "type": "int", + "value": [ + "32" + ] + }, + "channels_last": { + "type": "int", + "value": [ + "1" + ] + }, + "activation": { + "type": "int", + "value": [ + "0" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_25-/decoder/up_blocks.2/resnets.2/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_21.out12_21.out4_0_20" + ], + "const_args": [ + "NhwcConv_25_weight_NHWC" + ], + "out_args": [ + "NhwcConv_25_out-/decoder/up_blocks.2/resnets.2/conv1/Conv_output_0.out0_1_25" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "3", + "3", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_22", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_25_out-/decoder/up_blocks.2/resnets.2/conv1/Conv_output_0.out0_1_25" + ], + "const_args": [ + "GroupNorm_22_wts_4_0_21" + ], + "out_args": [ + "GroupNorm_22.out12_22.out4_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": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_26-/decoder/up_blocks.2/resnets.2/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_22.out12_22.out4_0_21" + ], + "const_args": [ + "NhwcConv_26_weight_NHWC" + ], + "out_args": [ + "NhwcConv_26_out-/decoder/up_blocks.2/resnets.2/conv2/Conv_output_0.out0_1_26" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "3", + "3", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/decoder/up_blocks.2/resnets.2/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/up_blocks.2/resnets.1/Add.out_2_1_10", + "NhwcConv_26_out-/decoder/up_blocks.2/resnets.2/conv2/Conv_output_0.out0_1_26" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.2/resnets.2/Add.out_2_1_11" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "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_2_1_11" + ], + "const_args": [ + "/decoder/up_blocks.2/upsamplers.0/Resize.weights7_2" + ], + "out_args": [ + "/decoder/up_blocks.2/upsamplers.0/Resize_output_0.nhwc5_2.out_7_2" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "256", + "256", + "256" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_27-/decoder/up_blocks.2/upsamplers.0/conv/Conv", + "type": "SDConv", + "in_args": [ + "/decoder/up_blocks.2/upsamplers.0/Resize_output_0.nhwc5_2.out_7_2" + ], + "const_args": [ + "NhwcConv_27_weight_NHWC" + ], + "out_args": [ + "NhwcConv_27_out-/decoder/up_blocks.2/upsamplers.0/conv/Conv_output_0.out0_1_27" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "256", + "3", + "3", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_23", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_27_out-/decoder/up_blocks.2/upsamplers.0/conv/Conv_output_0.out0_1_27" + ], + "const_args": [ + "GroupNorm_23_wts_4_0_22" + ], + "out_args": [ + "GroupNorm_23.out12_23.out4_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": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "512" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_28-/decoder/up_blocks.3/resnets.0/conv_shortcut/Conv", + "type": "SDConv", + "in_args": [ + "NhwcConv_27_out-/decoder/up_blocks.2/upsamplers.0/conv/Conv_output_0.out0_1_27" + ], + "const_args": [ + "NhwcConv_28_weight_NHWC" + ], + "out_args": [ + "NhwcConv_28_out-/decoder/up_blocks.3/resnets.0/conv_shortcut/Conv_output_0.out0_1_28" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "pads": { + "type": "int", + "value": [ + "0", + "0", + "0", + "0" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128", + "1", + "1", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "NhwcConv_29-/decoder/up_blocks.3/resnets.0/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_23.out12_23.out4_0_22" + ], + "const_args": [ + "NhwcConv_29_weight_NHWC" + ], + "out_args": [ + "NhwcConv_29_out-/decoder/up_blocks.3/resnets.0/conv1/Conv_output_0.out0_1_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": "int", + "value": [ + "1", + "512", + "512", + "256" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128", + "3", + "3", + "256" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_24", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_29_out-/decoder/up_blocks.3/resnets.0/conv1/Conv_output_0.out0_1_29" + ], + "const_args": [ + "GroupNorm_24_wts_4_0_23" + ], + "out_args": [ + "GroupNorm_24.out12_24.out4_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": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_30-/decoder/up_blocks.3/resnets.0/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_24.out12_24.out4_0_23" + ], + "const_args": [ + "NhwcConv_30_weight_NHWC" + ], + "out_args": [ + "NhwcConv_30_out-/decoder/up_blocks.3/resnets.0/conv2/Conv_output_0.out0_1_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": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128", + "3", + "3", + "128" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/decoder/up_blocks.3/resnets.0/Add", + "type": "SDAdd", + "in_args": [ + "NhwcConv_28_out-/decoder/up_blocks.3/resnets.0/conv_shortcut/Conv_output_0.out0_1_28", + "NhwcConv_30_out-/decoder/up_blocks.3/resnets.0/conv2/Conv_output_0.out0_1_30" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.3/resnets.0/Add.out_2_1_12" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "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_2_1_12" + ], + "const_args": [ + "GroupNorm_25_wts_4_0_24" + ], + "out_args": [ + "GroupNorm_25.out12_25.out4_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": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_31-/decoder/up_blocks.3/resnets.1/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_25.out12_25.out4_0_24" + ], + "const_args": [ + "NhwcConv_31_weight_NHWC" + ], + "out_args": [ + "NhwcConv_31_out-/decoder/up_blocks.3/resnets.1/conv1/Conv_output_0.out0_1_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": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128", + "3", + "3", + "128" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_26", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_31_out-/decoder/up_blocks.3/resnets.1/conv1/Conv_output_0.out0_1_31" + ], + "const_args": [ + "GroupNorm_26_wts_4_0_25" + ], + "out_args": [ + "GroupNorm_26.out12_26.out4_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": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_32-/decoder/up_blocks.3/resnets.1/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_26.out12_26.out4_0_25" + ], + "const_args": [ + "NhwcConv_32_weight_NHWC" + ], + "out_args": [ + "NhwcConv_32_out-/decoder/up_blocks.3/resnets.1/conv2/Conv_output_0.out0_1_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": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128", + "3", + "3", + "128" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/decoder/up_blocks.3/resnets.1/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/up_blocks.3/resnets.0/Add.out_2_1_12", + "NhwcConv_32_out-/decoder/up_blocks.3/resnets.1/conv2/Conv_output_0.out0_1_32" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.3/resnets.1/Add.out_2_1_13" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "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_2_1_13" + ], + "const_args": [ + "GroupNorm_27_wts_4_0_26" + ], + "out_args": [ + "GroupNorm_27.out12_27.out4_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": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_33-/decoder/up_blocks.3/resnets.2/conv1/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_27.out12_27.out4_0_26" + ], + "const_args": [ + "NhwcConv_33_weight_NHWC" + ], + "out_args": [ + "NhwcConv_33_out-/decoder/up_blocks.3/resnets.2/conv1/Conv_output_0.out0_1_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": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128", + "3", + "3", + "128" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "GroupNorm_28", + "type": "SDGroupNorm", + "in_args": [ + "NhwcConv_33_out-/decoder/up_blocks.3/resnets.2/conv1/Conv_output_0.out0_1_33" + ], + "const_args": [ + "GroupNorm_28_wts_4_0_27" + ], + "out_args": [ + "GroupNorm_28.out12_28.out4_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": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_34-/decoder/up_blocks.3/resnets.2/conv2/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_28.out12_28.out4_0_27" + ], + "const_args": [ + "NhwcConv_34_weight_NHWC" + ], + "out_args": [ + "NhwcConv_34_out-/decoder/up_blocks.3/resnets.2/conv2/Conv_output_0.out0_1_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": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "128", + "3", + "3", + "128" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + }, + { + "name": "/decoder/up_blocks.3/resnets.2/Add", + "type": "SDAdd", + "in_args": [ + "/decoder/up_blocks.3/resnets.1/Add.out_2_1_13", + "NhwcConv_34_out-/decoder/up_blocks.3/resnets.2/conv2/Conv_output_0.out0_1_34" + ], + "const_args": [], + "out_args": [ + "/decoder/up_blocks.3/resnets.2/Add.out_2_1_14" + ], + "attrs": { + "a_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "b_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "c_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "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_2_1_14" + ], + "const_args": [ + "GroupNorm_29_wts_4_0_28" + ], + "out_args": [ + "GroupNorm_29.out12_29.out4_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": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "wts_shape": { + "type": "int", + "value": [ + "256" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfloat16" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "gamma_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "beta_shape": { + "type": "int", + "value": [ + "128" + ] + }, + "nonlinear": { + "type": "str", + "value": [ + "Silu" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + } + } + }, + { + "name": "NhwcConv_35-/decoder/conv_out/Conv", + "type": "SDConv", + "in_args": [ + "GroupNorm_29.out12_29.out4_0_28" + ], + "const_args": [ + "NhwcConv_35_weight_NHWC" + ], + "out_args": [ + "NhwcConv_35_out-sample.out0_1_35" + ], + "attrs": { + "auto_pad": { + "type": "str", + "value": [ + "NOTSET" + ] + }, + "dilations": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "group": { + "type": "int", + "value": [ + "1" + ] + }, + "kernel_shape": { + "type": "int", + "value": [ + "3", + "3" + ] + }, + "pads": { + "type": "int", + "value": [ + "1", + "1", + "1", + "1" + ] + }, + "strides": { + "type": "int", + "value": [ + "1", + "1" + ] + }, + "input_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "128" + ] + }, + "output_shape": { + "type": "int", + "value": [ + "1", + "512", + "512", + "3" + ] + }, + "weight_shape": { + "type": "int", + "value": [ + "3", + "3", + "3", + "128" + ] + }, + "out_dtypes": { + "type": "str", + "value": [ + "bfloat16" + ] + }, + "ctrl_packet": { + "type": "int", + "value": [ + "1" + ] + }, + "in_dtypes": { + "type": "str", + "value": [ + "bfloat16", + "bfp16ebs8", + "float" + ] + } + } + } + ], + "fused_tensors": { + "in": { + "buffer_size": 32768, + "xrt_arg_id": 0, + "packed_tensors": [ + "Transpose_0_out-latents.out0_1_0" + ] + }, + "out": { + "buffer_size": 1572864, + "xrt_arg_id": 1, + "packed_tensors": [ + "NhwcConv_35_out-sample.out0_1_35" + ] + }, + "scratch": { + "buffer_size": 2604695552, + "xrt_arg_id": 2, + "packed_tensors": [ + "NhwcConv_0_out-/post_quant_conv/Conv_output_0.out0_1_0", + "NhwcConv_1_out-/decoder/conv_in/Conv_output_0.out0_1_1", + "GroupNorm_0.out12_0.out4_0_0", + "NhwcConv_2_out-/decoder/mid_block/resnets.0/conv1/Conv_output_0.out0_1_2", + "GroupNorm_1.out12_1.out4_0_1", + "NhwcConv_3_out-/decoder/mid_block/resnets.0/conv2/Conv_output_0.out0_1_3", + "/decoder/mid_block/resnets.0/Add.out_2_1_0", + "/decoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out4_1_0", + "/decoder/mid_block/attentions.0/Reshape_2_output_0_3d.out1_3_0", + "/decoder/mid_block/attentions.0/Reshape_1_output_0_3d.out1_3_1", + "/decoder/mid_block/attentions.0/Reshape_3_output_0_3d.out1_3_2", + "/decoder/mid_block/attentions.0/Transpose_5_output_0_3d.out10_0", + "/decoder/mid_block/attentions.0/Reshape_5_output_0.nhwc6_0.out1_0_0", + "/decoder/mid_block/attentions.0/Add_output_0_NHWC.out_2_1_1", + "GroupNorm_3.out12_3.out4_0_2", + "NhwcConv_4_out-/decoder/mid_block/resnets.1/conv1/Conv_output_0.out0_1_4", + "GroupNorm_4.out12_4.out4_0_3", + "NhwcConv_5_out-/decoder/mid_block/resnets.1/conv2/Conv_output_0.out0_1_5", + "/decoder/mid_block/resnets.1/Add_output_0_NHWC.out_2_1_2", + "GroupNorm_5.out12_5.out4_0_4", + "NhwcConv_6_out-/decoder/up_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_6", + "GroupNorm_6.out12_6.out4_0_5", + "NhwcConv_7_out-/decoder/up_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_7", + "/decoder/up_blocks.0/resnets.0/Add_output_0_NHWC.out_2_1_3", + "GroupNorm_7.out12_7.out4_0_6", + "NhwcConv_8_out-/decoder/up_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_8", + "GroupNorm_8.out12_8.out4_0_7", + "NhwcConv_9_out-/decoder/up_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_9", + "/decoder/up_blocks.0/resnets.1/Add_output_0_NHWC.out_2_1_4", + "GroupNorm_9.out12_9.out4_0_8", + "NhwcConv_10_out-/decoder/up_blocks.0/resnets.2/conv1/Conv_output_0.out0_1_10", + "GroupNorm_10.out12_10.out4_0_9", + "NhwcConv_11_out-/decoder/up_blocks.0/resnets.2/conv2/Conv_output_0.out0_1_11", + "/decoder/up_blocks.0/resnets.2/Add_output_0.nhwc5_0.out_2_1_5", + "/decoder/up_blocks.0/upsamplers.0/Resize_output_0.nhwc5_0.out_7_0", + "NhwcConv_12_out-/decoder/up_blocks.0/upsamplers.0/conv/Conv_output_0.out0_1_12", + "GroupNorm_11.out12_11.out4_0_10", + "NhwcConv_13_out-/decoder/up_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_13", + "GroupNorm_12.out12_12.out4_0_11", + "NhwcConv_14_out-/decoder/up_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_14", + "/decoder/up_blocks.1/resnets.0/Add.out_2_1_6", + "GroupNorm_13.out12_13.out4_0_12", + "NhwcConv_15_out-/decoder/up_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_15", + "GroupNorm_14.out12_14.out4_0_13", + "NhwcConv_16_out-/decoder/up_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_16", + "/decoder/up_blocks.1/resnets.1/Add.out_2_1_7", + "GroupNorm_15.out12_15.out4_0_14", + "NhwcConv_17_out-/decoder/up_blocks.1/resnets.2/conv1/Conv_output_0.out0_1_17", + "GroupNorm_16.out12_16.out4_0_15", + "NhwcConv_18_out-/decoder/up_blocks.1/resnets.2/conv2/Conv_output_0.out0_1_18", + "/decoder/up_blocks.1/resnets.2/Add.out_2_1_8", + "/decoder/up_blocks.1/upsamplers.0/Resize_output_0.nhwc5_1.out_7_1", + "NhwcConv_19_out-/decoder/up_blocks.1/upsamplers.0/conv/Conv_output_0.out0_1_19", + "GroupNorm_17.out12_17.out4_0_16", + "NhwcConv_20_out-/decoder/up_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_20", + "NhwcConv_21_out-/decoder/up_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_21", + "GroupNorm_18.out12_18.out4_0_17", + "NhwcConv_22_out-/decoder/up_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_22", + "/decoder/up_blocks.2/resnets.0/Add.out_2_1_9", + "GroupNorm_19.out12_19.out4_0_18", + "NhwcConv_23_out-/decoder/up_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_23", + "GroupNorm_20.out12_20.out4_0_19", + "NhwcConv_24_out-/decoder/up_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_24", + "/decoder/up_blocks.2/resnets.1/Add.out_2_1_10", + "GroupNorm_21.out12_21.out4_0_20", + "NhwcConv_25_out-/decoder/up_blocks.2/resnets.2/conv1/Conv_output_0.out0_1_25", + "GroupNorm_22.out12_22.out4_0_21", + "NhwcConv_26_out-/decoder/up_blocks.2/resnets.2/conv2/Conv_output_0.out0_1_26", + "/decoder/up_blocks.2/resnets.2/Add.out_2_1_11", + "/decoder/up_blocks.2/upsamplers.0/Resize_output_0.nhwc5_2.out_7_2", + "NhwcConv_27_out-/decoder/up_blocks.2/upsamplers.0/conv/Conv_output_0.out0_1_27", + "GroupNorm_23.out12_23.out4_0_22", + "NhwcConv_28_out-/decoder/up_blocks.3/resnets.0/conv_shortcut/Conv_output_0.out0_1_28", + "NhwcConv_29_out-/decoder/up_blocks.3/resnets.0/conv1/Conv_output_0.out0_1_29", + "GroupNorm_24.out12_24.out4_0_23", + "NhwcConv_30_out-/decoder/up_blocks.3/resnets.0/conv2/Conv_output_0.out0_1_30", + "/decoder/up_blocks.3/resnets.0/Add.out_2_1_12", + "GroupNorm_25.out12_25.out4_0_24", + "NhwcConv_31_out-/decoder/up_blocks.3/resnets.1/conv1/Conv_output_0.out0_1_31", + "GroupNorm_26.out12_26.out4_0_25", + "NhwcConv_32_out-/decoder/up_blocks.3/resnets.1/conv2/Conv_output_0.out0_1_32", + "/decoder/up_blocks.3/resnets.1/Add.out_2_1_13", + "GroupNorm_27.out12_27.out4_0_26", + "NhwcConv_33_out-/decoder/up_blocks.3/resnets.2/conv1/Conv_output_0.out0_1_33", + "GroupNorm_28.out12_28.out4_0_27", + "NhwcConv_34_out-/decoder/up_blocks.3/resnets.2/conv2/Conv_output_0.out0_1_34", + "/decoder/up_blocks.3/resnets.2/Add.out_2_1_14", + "GroupNorm_29.out12_29.out4_0_28" + ] + }, + "const": { + "buffer_size": 61089680, + "xrt_arg_id": 3, + "packed_tensors": [ + "NhwcConv_0_weight_NHWC", + "NhwcConv_1_weight_NHWC", + "GroupNorm_0_wts_4_0_0", + "NhwcConv_2_weight_NHWC", + "GroupNorm_1_wts_4_0_1", + "NhwcConv_3_weight_NHWC", + "GroupNorm_2_wts_4_1_0", + "onnx::MatMul_918_2_0_0", + "onnx::MatMul_917_2_0_1", + "onnx::MatMul_919_2_0_2", + "/decoder/mid_block/attentions.0/MatMul/MatMulScaleFusion/_FusedActivationmha_3_1_0_mask.10_0", + "onnx::MatMul_927_2_0_3", + "GroupNorm_3_wts_4_0_2", + "NhwcConv_4_weight_NHWC", + "GroupNorm_4_wts_4_0_3", + "NhwcConv_5_weight_NHWC", + "GroupNorm_5_wts_4_0_4", + "NhwcConv_6_weight_NHWC", + "GroupNorm_6_wts_4_0_5", + "NhwcConv_7_weight_NHWC", + "GroupNorm_7_wts_4_0_6", + "NhwcConv_8_weight_NHWC", + "GroupNorm_8_wts_4_0_7", + "NhwcConv_9_weight_NHWC", + "GroupNorm_9_wts_4_0_8", + "NhwcConv_10_weight_NHWC", + "GroupNorm_10_wts_4_0_9", + "NhwcConv_11_weight_NHWC", + "/decoder/up_blocks.0/upsamplers.0/Resize.weights7_0", + "NhwcConv_12_weight_NHWC", + "GroupNorm_11_wts_4_0_10", + "NhwcConv_13_weight_NHWC", + "GroupNorm_12_wts_4_0_11", + "NhwcConv_14_weight_NHWC", + "GroupNorm_13_wts_4_0_12", + "NhwcConv_15_weight_NHWC", + "GroupNorm_14_wts_4_0_13", + "NhwcConv_16_weight_NHWC", + "GroupNorm_15_wts_4_0_14", + "NhwcConv_17_weight_NHWC", + "GroupNorm_16_wts_4_0_15", + "NhwcConv_18_weight_NHWC", + "/decoder/up_blocks.1/upsamplers.0/Resize.weights7_1", + "NhwcConv_19_weight_NHWC", + "GroupNorm_17_wts_4_0_16", + "NhwcConv_20_weight_NHWC", + "NhwcConv_21_weight_NHWC", + "GroupNorm_18_wts_4_0_17", + "NhwcConv_22_weight_NHWC", + "GroupNorm_19_wts_4_0_18", + "NhwcConv_23_weight_NHWC", + "GroupNorm_20_wts_4_0_19", + "NhwcConv_24_weight_NHWC", + "GroupNorm_21_wts_4_0_20", + "NhwcConv_25_weight_NHWC", + "GroupNorm_22_wts_4_0_21", + "NhwcConv_26_weight_NHWC", + "/decoder/up_blocks.2/upsamplers.0/Resize.weights7_2", + "NhwcConv_27_weight_NHWC", + "GroupNorm_23_wts_4_0_22", + "NhwcConv_28_weight_NHWC", + "NhwcConv_29_weight_NHWC", + "GroupNorm_24_wts_4_0_23", + "NhwcConv_30_weight_NHWC", + "GroupNorm_25_wts_4_0_24", + "NhwcConv_31_weight_NHWC", + "GroupNorm_26_wts_4_0_25", + "NhwcConv_32_weight_NHWC", + "GroupNorm_27_wts_4_0_26", + "NhwcConv_33_weight_NHWC", + "GroupNorm_28_wts_4_0_27", + "NhwcConv_34_weight_NHWC", + "GroupNorm_29_wts_4_0_28", + "NhwcConv_35_weight_NHWC" + ] + }, + "super_instr": { + "buffer_size": 0, + "xrt_arg_id": 4, + "packed_tensors": [] + } + }, + "tensor_map": { + "Transpose_0_out-latents.out0_1_0": { + "packed_buffer_label": "in", + "xrt_arg_id": 0, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 4 + ], + "size_in_bytes": 32768, + "op_tensor_size": 32768, + "offset": 0 + }, + "NhwcConv_35_out-sample.out0_1_35": { + "packed_buffer_label": "out", + "xrt_arg_id": 1, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 3 + ], + "size_in_bytes": 1572864, + "op_tensor_size": 1572864, + "offset": 0 + }, + "NhwcConv_0_out-/post_quant_conv/Conv_output_0.out0_1_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 4 + ], + "size_in_bytes": 32768, + "op_tensor_size": 32768, + "offset": 0 + }, + "NhwcConv_1_out-/decoder/conv_in/Conv_output_0.out0_1_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 32768 + }, + "GroupNorm_0.out12_0.out4_0_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 4227072 + }, + "NhwcConv_2_out-/decoder/mid_block/resnets.0/conv1/Conv_output_0.out0_1_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 8421376 + }, + "GroupNorm_1.out12_1.out4_0_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 12615680 + }, + "NhwcConv_3_out-/decoder/mid_block/resnets.0/conv2/Conv_output_0.out0_1_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 16809984 + }, + "/decoder/mid_block/resnets.0/Add.out_2_1_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 21004288 + }, + "/decoder/mid_block/attentions.0/group_norm/Add_output_0_4d_NHWC.out4_1_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 4096, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 25198592 + }, + "/decoder/mid_block/attentions.0/Reshape_2_output_0_3d.out1_3_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 4096, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 29392896 + }, + "/decoder/mid_block/attentions.0/Reshape_1_output_0_3d.out1_3_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 4096, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 33587200 + }, + "/decoder/mid_block/attentions.0/Reshape_3_output_0_3d.out1_3_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 4096, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 37781504 + }, + "/decoder/mid_block/attentions.0/Transpose_5_output_0_3d.out10_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 4096, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 41975808 + }, + "/decoder/mid_block/attentions.0/Reshape_5_output_0.nhwc6_0.out1_0_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 46170112 + }, + "/decoder/mid_block/attentions.0/Add_output_0_NHWC.out_2_1_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 50364416 + }, + "GroupNorm_3.out12_3.out4_0_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 54558720 + }, + "NhwcConv_4_out-/decoder/mid_block/resnets.1/conv1/Conv_output_0.out0_1_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 58753024 + }, + "GroupNorm_4.out12_4.out4_0_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 62947328 + }, + "NhwcConv_5_out-/decoder/mid_block/resnets.1/conv2/Conv_output_0.out0_1_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 67141632 + }, + "/decoder/mid_block/resnets.1/Add_output_0_NHWC.out_2_1_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 71335936 + }, + "GroupNorm_5.out12_5.out4_0_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 75530240 + }, + "NhwcConv_6_out-/decoder/up_blocks.0/resnets.0/conv1/Conv_output_0.out0_1_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 79724544 + }, + "GroupNorm_6.out12_6.out4_0_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 83918848 + }, + "NhwcConv_7_out-/decoder/up_blocks.0/resnets.0/conv2/Conv_output_0.out0_1_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 88113152 + }, + "/decoder/up_blocks.0/resnets.0/Add_output_0_NHWC.out_2_1_3": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 92307456 + }, + "GroupNorm_7.out12_7.out4_0_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 96501760 + }, + "NhwcConv_8_out-/decoder/up_blocks.0/resnets.1/conv1/Conv_output_0.out0_1_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 100696064 + }, + "GroupNorm_8.out12_8.out4_0_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 104890368 + }, + "NhwcConv_9_out-/decoder/up_blocks.0/resnets.1/conv2/Conv_output_0.out0_1_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 109084672 + }, + "/decoder/up_blocks.0/resnets.1/Add_output_0_NHWC.out_2_1_4": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 113278976 + }, + "GroupNorm_9.out12_9.out4_0_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 117473280 + }, + "NhwcConv_10_out-/decoder/up_blocks.0/resnets.2/conv1/Conv_output_0.out0_1_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 121667584 + }, + "GroupNorm_10.out12_10.out4_0_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 125861888 + }, + "NhwcConv_11_out-/decoder/up_blocks.0/resnets.2/conv2/Conv_output_0.out0_1_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 130056192 + }, + "/decoder/up_blocks.0/resnets.2/Add_output_0.nhwc5_0.out_2_1_5": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 64, + 64, + 512 + ], + "size_in_bytes": 4194304, + "op_tensor_size": 4194304, + "offset": 134250496 + }, + "/decoder/up_blocks.0/upsamplers.0/Resize_output_0.nhwc5_0.out_7_0": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 138444800 + }, + "NhwcConv_12_out-/decoder/up_blocks.0/upsamplers.0/conv/Conv_output_0.out0_1_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 155222016 + }, + "GroupNorm_11.out12_11.out4_0_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 171999232 + }, + "NhwcConv_13_out-/decoder/up_blocks.1/resnets.0/conv1/Conv_output_0.out0_1_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 188776448 + }, + "GroupNorm_12.out12_12.out4_0_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 205553664 + }, + "NhwcConv_14_out-/decoder/up_blocks.1/resnets.0/conv2/Conv_output_0.out0_1_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 222330880 + }, + "/decoder/up_blocks.1/resnets.0/Add.out_2_1_6": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 239108096 + }, + "GroupNorm_13.out12_13.out4_0_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 255885312 + }, + "NhwcConv_15_out-/decoder/up_blocks.1/resnets.1/conv1/Conv_output_0.out0_1_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 272662528 + }, + "GroupNorm_14.out12_14.out4_0_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 289439744 + }, + "NhwcConv_16_out-/decoder/up_blocks.1/resnets.1/conv2/Conv_output_0.out0_1_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 306216960 + }, + "/decoder/up_blocks.1/resnets.1/Add.out_2_1_7": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 322994176 + }, + "GroupNorm_15.out12_15.out4_0_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 339771392 + }, + "NhwcConv_17_out-/decoder/up_blocks.1/resnets.2/conv1/Conv_output_0.out0_1_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 356548608 + }, + "GroupNorm_16.out12_16.out4_0_15": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 373325824 + }, + "NhwcConv_18_out-/decoder/up_blocks.1/resnets.2/conv2/Conv_output_0.out0_1_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 390103040 + }, + "/decoder/up_blocks.1/resnets.2/Add.out_2_1_8": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 128, + 128, + 512 + ], + "size_in_bytes": 16777216, + "op_tensor_size": 16777216, + "offset": 406880256 + }, + "/decoder/up_blocks.1/upsamplers.0/Resize_output_0.nhwc5_1.out_7_1": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 512 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 423657472 + }, + "NhwcConv_19_out-/decoder/up_blocks.1/upsamplers.0/conv/Conv_output_0.out0_1_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 512 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 490766336 + }, + "GroupNorm_17.out12_17.out4_0_16": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 512 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 557875200 + }, + "NhwcConv_20_out-/decoder/up_blocks.2/resnets.0/conv_shortcut/Conv_output_0.out0_1_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 256 + ], + "size_in_bytes": 33554432, + "op_tensor_size": 33554432, + "offset": 624984064 + }, + "NhwcConv_21_out-/decoder/up_blocks.2/resnets.0/conv1/Conv_output_0.out0_1_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 256 + ], + "size_in_bytes": 33554432, + "op_tensor_size": 33554432, + "offset": 658538496 + }, + "GroupNorm_18.out12_18.out4_0_17": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 256 + ], + "size_in_bytes": 33554432, + "op_tensor_size": 33554432, + "offset": 692092928 + }, + "NhwcConv_22_out-/decoder/up_blocks.2/resnets.0/conv2/Conv_output_0.out0_1_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 256 + ], + "size_in_bytes": 33554432, + "op_tensor_size": 33554432, + "offset": 725647360 + }, + "/decoder/up_blocks.2/resnets.0/Add.out_2_1_9": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 256 + ], + "size_in_bytes": 33554432, + "op_tensor_size": 33554432, + "offset": 759201792 + }, + "GroupNorm_19.out12_19.out4_0_18": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 256 + ], + "size_in_bytes": 33554432, + "op_tensor_size": 33554432, + "offset": 792756224 + }, + "NhwcConv_23_out-/decoder/up_blocks.2/resnets.1/conv1/Conv_output_0.out0_1_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 256 + ], + "size_in_bytes": 33554432, + "op_tensor_size": 33554432, + "offset": 826310656 + }, + "GroupNorm_20.out12_20.out4_0_19": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 256 + ], + "size_in_bytes": 33554432, + "op_tensor_size": 33554432, + "offset": 859865088 + }, + "NhwcConv_24_out-/decoder/up_blocks.2/resnets.1/conv2/Conv_output_0.out0_1_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 256 + ], + "size_in_bytes": 33554432, + "op_tensor_size": 33554432, + "offset": 893419520 + }, + "/decoder/up_blocks.2/resnets.1/Add.out_2_1_10": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 256 + ], + "size_in_bytes": 33554432, + "op_tensor_size": 33554432, + "offset": 926973952 + }, + "GroupNorm_21.out12_21.out4_0_20": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 256 + ], + "size_in_bytes": 33554432, + "op_tensor_size": 33554432, + "offset": 960528384 + }, + "NhwcConv_25_out-/decoder/up_blocks.2/resnets.2/conv1/Conv_output_0.out0_1_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 256 + ], + "size_in_bytes": 33554432, + "op_tensor_size": 33554432, + "offset": 994082816 + }, + "GroupNorm_22.out12_22.out4_0_21": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 256 + ], + "size_in_bytes": 33554432, + "op_tensor_size": 33554432, + "offset": 1027637248 + }, + "NhwcConv_26_out-/decoder/up_blocks.2/resnets.2/conv2/Conv_output_0.out0_1_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 256 + ], + "size_in_bytes": 33554432, + "op_tensor_size": 33554432, + "offset": 1061191680 + }, + "/decoder/up_blocks.2/resnets.2/Add.out_2_1_11": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 256, + 256, + 256 + ], + "size_in_bytes": 33554432, + "op_tensor_size": 33554432, + "offset": 1094746112 + }, + "/decoder/up_blocks.2/upsamplers.0/Resize_output_0.nhwc5_2.out_7_2": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 256 + ], + "size_in_bytes": 134217728, + "op_tensor_size": 134217728, + "offset": 1128300544 + }, + "NhwcConv_27_out-/decoder/up_blocks.2/upsamplers.0/conv/Conv_output_0.out0_1_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 256 + ], + "size_in_bytes": 134217728, + "op_tensor_size": 134217728, + "offset": 1262518272 + }, + "GroupNorm_23.out12_23.out4_0_22": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 256 + ], + "size_in_bytes": 134217728, + "op_tensor_size": 134217728, + "offset": 1396736000 + }, + "NhwcConv_28_out-/decoder/up_blocks.3/resnets.0/conv_shortcut/Conv_output_0.out0_1_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 128 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 1530953728 + }, + "NhwcConv_29_out-/decoder/up_blocks.3/resnets.0/conv1/Conv_output_0.out0_1_29": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 128 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 1598062592 + }, + "GroupNorm_24.out12_24.out4_0_23": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 128 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 1665171456 + }, + "NhwcConv_30_out-/decoder/up_blocks.3/resnets.0/conv2/Conv_output_0.out0_1_30": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 128 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 1732280320 + }, + "/decoder/up_blocks.3/resnets.0/Add.out_2_1_12": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 128 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 1799389184 + }, + "GroupNorm_25.out12_25.out4_0_24": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 128 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 1866498048 + }, + "NhwcConv_31_out-/decoder/up_blocks.3/resnets.1/conv1/Conv_output_0.out0_1_31": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 128 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 1933606912 + }, + "GroupNorm_26.out12_26.out4_0_25": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 128 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 2000715776 + }, + "NhwcConv_32_out-/decoder/up_blocks.3/resnets.1/conv2/Conv_output_0.out0_1_32": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 128 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 2067824640 + }, + "/decoder/up_blocks.3/resnets.1/Add.out_2_1_13": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 128 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 2134933504 + }, + "GroupNorm_27.out12_27.out4_0_26": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 128 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 2202042368 + }, + "NhwcConv_33_out-/decoder/up_blocks.3/resnets.2/conv1/Conv_output_0.out0_1_33": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 128 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 2269151232 + }, + "GroupNorm_28.out12_28.out4_0_27": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 128 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 2336260096 + }, + "NhwcConv_34_out-/decoder/up_blocks.3/resnets.2/conv2/Conv_output_0.out0_1_34": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 128 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 2403368960 + }, + "/decoder/up_blocks.3/resnets.2/Add.out_2_1_14": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 128 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 2470477824 + }, + "GroupNorm_29.out12_29.out4_0_28": { + "packed_buffer_label": "scratch", + "xrt_arg_id": 2, + "dtype": "bfloat16", + "shape": [ + 1, + 512, + 512, + 128 + ], + "size_in_bytes": 67108864, + "op_tensor_size": 67108864, + "offset": 2537586688 + }, + "NhwcConv_0_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 400 + ], + "size_in_bytes": 400, + "op_tensor_size": 400, + "offset": 0, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_0.const", + "file_size": 400 + }, + "NhwcConv_1_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 49664 + ], + "size_in_bytes": 49664, + "op_tensor_size": 49664, + "offset": 400, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_1.const", + "file_size": 49664 + }, + "GroupNorm_0_wts_4_0_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 50064, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_2.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": 52112, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_3.const", + "file_size": 2916352 + }, + "GroupNorm_1_wts_4_0_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 2968464, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_4.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": 2970512, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_5.const", + "file_size": 2916352 + }, + "GroupNorm_2_wts_4_1_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 5886864, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_6.const", + "file_size": 2048 + }, + "onnx::MatMul_918_2_0_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 299008 + ], + "size_in_bytes": 299008, + "op_tensor_size": 299008, + "offset": 5888912, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_7.const", + "file_size": 299008 + }, + "onnx::MatMul_917_2_0_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 299008 + ], + "size_in_bytes": 299008, + "op_tensor_size": 299008, + "offset": 6187920, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_8.const", + "file_size": 299008 + }, + "onnx::MatMul_919_2_0_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 299008 + ], + "size_in_bytes": 299008, + "op_tensor_size": 299008, + "offset": 6486928, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_9.const", + "file_size": 299008 + }, + "/decoder/mid_block/attentions.0/MatMul/MatMulScaleFusion/_FusedActivationmha_3_1_0_mask.10_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 4096 + ], + "size_in_bytes": 8192, + "op_tensor_size": 8192, + "offset": 6785936, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_10.const", + "file_size": 8192 + }, + "onnx::MatMul_927_2_0_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 299008 + ], + "size_in_bytes": 299008, + "op_tensor_size": 299008, + "offset": 6794128, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_11.const", + "file_size": 299008 + }, + "GroupNorm_3_wts_4_0_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 7093136, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_12.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": 7095184, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_13.const", + "file_size": 2916352 + }, + "GroupNorm_4_wts_4_0_3": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 10011536, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_14.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": 10013584, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_15.const", + "file_size": 2916352 + }, + "GroupNorm_5_wts_4_0_4": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 12929936, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_16.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": 12931984, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_17.const", + "file_size": 2916352 + }, + "GroupNorm_6_wts_4_0_5": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 15848336, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_18.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": 15850384, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_19.const", + "file_size": 2916352 + }, + "GroupNorm_7_wts_4_0_6": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 18766736, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_20.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": 18768784, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_21.const", + "file_size": 2916352 + }, + "GroupNorm_8_wts_4_0_7": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 21685136, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_22.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": 21687184, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_23.const", + "file_size": 2916352 + }, + "GroupNorm_9_wts_4_0_8": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 24603536, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_24.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": 24605584, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_25.const", + "file_size": 2916352 + }, + "GroupNorm_10_wts_4_0_9": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 27521936, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_26.const", + "file_size": 2048 + }, + "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": 27523984, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_27.const", + "file_size": 2916352 + }, + "/decoder/up_blocks.0/upsamplers.0/Resize.weights7_0": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "offset": 30440336, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_28.const", + "file_size": 256 + }, + "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": 30440592, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_29.const", + "file_size": 2916352 + }, + "GroupNorm_11_wts_4_0_10": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 33356944, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_30.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": 33358992, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_31.const", + "file_size": 2916352 + }, + "GroupNorm_12_wts_4_0_11": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 36275344, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_32.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": 36277392, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_33.const", + "file_size": 2916352 + }, + "GroupNorm_13_wts_4_0_12": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 39193744, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_34.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": 39195792, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_35.const", + "file_size": 2916352 + }, + "GroupNorm_14_wts_4_0_13": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 42112144, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_36.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": 42114192, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_37.const", + "file_size": 2916352 + }, + "GroupNorm_15_wts_4_0_14": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 45030544, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_38.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": 45032592, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_39.const", + "file_size": 2916352 + }, + "GroupNorm_16_wts_4_0_15": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 47948944, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_40.const", + "file_size": 2048 + }, + "NhwcConv_18_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 47950992, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_41.const", + "file_size": 2916352 + }, + "/decoder/up_blocks.1/upsamplers.0/Resize.weights7_1": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "offset": 50867344, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_42.const", + "file_size": 256 + }, + "NhwcConv_19_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 2916352 + ], + "size_in_bytes": 2916352, + "op_tensor_size": 2916352, + "offset": 50867600, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_43.const", + "file_size": 2916352 + }, + "GroupNorm_17_wts_4_0_16": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 1024 + ], + "size_in_bytes": 2048, + "op_tensor_size": 2048, + "offset": 53783952, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_44.const", + "file_size": 2048 + }, + "NhwcConv_20_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 163840 + ], + "size_in_bytes": 163840, + "op_tensor_size": 163840, + "offset": 53786000, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_45.const", + "file_size": 163840 + }, + "NhwcConv_21_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 1458176 + ], + "size_in_bytes": 1458176, + "op_tensor_size": 1458176, + "offset": 53949840, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_46.const", + "file_size": 1458176 + }, + "GroupNorm_18_wts_4_0_17": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "offset": 55408016, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_47.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": 55409040, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_48.const", + "file_size": 729088 + }, + "GroupNorm_19_wts_4_0_18": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "offset": 56138128, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_49.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": 56139152, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_50.const", + "file_size": 729088 + }, + "GroupNorm_20_wts_4_0_19": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "offset": 56868240, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_51.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": 56869264, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_52.const", + "file_size": 729088 + }, + "GroupNorm_21_wts_4_0_20": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "offset": 57598352, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_53.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": 57599376, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_54.const", + "file_size": 729088 + }, + "GroupNorm_22_wts_4_0_21": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "offset": 58328464, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_55.const", + "file_size": 1024 + }, + "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": 58329488, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_56.const", + "file_size": 729088 + }, + "/decoder/up_blocks.2/upsamplers.0/Resize.weights7_2": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 128 + ], + "size_in_bytes": 256, + "op_tensor_size": 256, + "offset": 59058576, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_57.const", + "file_size": 256 + }, + "NhwcConv_27_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 729088 + ], + "size_in_bytes": 729088, + "op_tensor_size": 729088, + "offset": 59058832, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_58.const", + "file_size": 729088 + }, + "GroupNorm_23_wts_4_0_22": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 512 + ], + "size_in_bytes": 1024, + "op_tensor_size": 1024, + "offset": 59787920, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_59.const", + "file_size": 1024 + }, + "NhwcConv_28_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 40960 + ], + "size_in_bytes": 40960, + "op_tensor_size": 40960, + "offset": 59788944, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_60.const", + "file_size": 40960 + }, + "NhwcConv_29_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 364544 + ], + "size_in_bytes": 364544, + "op_tensor_size": 364544, + "offset": 59829904, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_61.const", + "file_size": 364544 + }, + "GroupNorm_24_wts_4_0_23": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "offset": 60194448, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_62.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": 60194960, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_63.const", + "file_size": 174080 + }, + "GroupNorm_25_wts_4_0_24": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "offset": 60369040, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_64.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": 60369552, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_65.const", + "file_size": 174080 + }, + "GroupNorm_26_wts_4_0_25": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "offset": 60543632, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_66.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": 60544144, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_67.const", + "file_size": 174080 + }, + "GroupNorm_27_wts_4_0_26": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "offset": 60718224, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_68.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": 60718736, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_69.const", + "file_size": 174080 + }, + "GroupNorm_28_wts_4_0_27": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "offset": 60892816, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_70.const", + "file_size": 512 + }, + "NhwcConv_34_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 174080 + ], + "size_in_bytes": 174080, + "op_tensor_size": 174080, + "offset": 60893328, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_71.const", + "file_size": 174080 + }, + "GroupNorm_29_wts_4_0_28": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfloat16", + "shape": [ + 256 + ], + "size_in_bytes": 512, + "op_tensor_size": 512, + "offset": 61067408, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_72.const", + "file_size": 512 + }, + "NhwcConv_35_weight_NHWC": { + "packed_buffer_label": "const", + "xrt_arg_id": 3, + "dtype": "bfp16ebs8", + "shape": [ + 21760 + ], + "size_in_bytes": 21760, + "op_tensor_size": 21760, + "offset": 61067920, + "file_name": ".cache/NhwcConv_0-post_quant_convConv_73.const", + "file_size": 21760 + } + }, + "aux_info": {} +} \ No newline at end of file