digraph optimistic_mid {
  rankdir=TB; nodesep=0.2; ranksep=0.35;
  node [fontname=monospace,fontsize=11];
  v0 [shape=box,label="v0 / high 0\n[1, 8, 128]\nF16 RowMajor\n64 owners, 1 replicas"];
  v1 [shape=box,label="v1 / high 1\n[1, 1, 128]\nF16 RowMajor\n64 owners, 1 replicas"];
  v2 [shape=box,label="v2 / high 2\n[1, 1, 128]\nF16 RowMajor\n64 owners, 1 replicas"];
  v3 [shape=box,label="v3 / high 4\n[128, 128]\nF16 RowMajor\n64 owners, 1 replicas"];
  v4 [shape=box,label="v4 / high 5\n[128, 128]\nF16 RowMajor\n64 owners, 1 replicas"];
  v5 [shape=box,label="v5 / high 3\n[1, 8, 128]\nF16 RowMajor\n64 owners, 1 replicas"];
  v6 [shape=box,label="v6 / high 3\n[1, 8, 128]\nF16 RowMajor\n64 owners, 16 replicas"];
  v7 [shape=box,label="v7 / high 3\n[1, 8, 128]\nF8F143 { scale_exponent: -4 }\nBlockMajor(TransposedMatrix { row_block: 32,\ncolumn_block: 16 })\n64 owners, 16 replicas"];
  v8 [shape=box,label="v8 / high 4\n[128, 128]\nF16 RowMajor\n64 owners, 1 replicas"];
  v9 [shape=box,label="v9 / high 4\n[128, 128]\nF8F143 { scale_exponent: -4 } Amp(TransposedLeft)\n64 owners, 1 replicas"];
  v10 [shape=box,label="v10 / high 6\n[1, 8, 128]\nF16 Amp(TransposedLeft)\n64 owners, 1 replicas"];
  v11 [shape=box,label="v11 / high 5\n[128, 128]\nF16 RowMajor\n64 owners, 1 replicas"];
  v12 [shape=box,label="v12 / high 5\n[128, 128]\nF8F143 { scale_exponent: -4 } Amp(TransposedLeft)\n64 owners, 1 replicas"];
  v13 [shape=box,label="v13 / high 7\n[1, 8, 128]\nF16 Amp(TransposedLeft)\n64 owners, 1 replicas"];
  v14 [shape=box,label="v14 / high 6\n[1, 8, 128]\nF16 RowMajor\n64 owners, 1 replicas"];
  v15 [shape=box,label="v15 / high 6\n[1, 8, 128]\nF16 RowMajor\n64 owners, 1 replicas"];
  v16 [shape=box,label="v16 / high 7\n[1, 8, 128]\nF16 RowMajor\n64 owners, 1 replicas"];
  v17 [shape=box,label="v17 / high 7\n[1, 8, 128]\nF16 RowMajor\n64 owners, 1 replicas"];
  s0 [label="LayerNorm\n3963 / 3963 cycles\n{}",color=black];
  v0 -> s0;
  v1 -> s0;
  v2 -> s0;
  s0 -> v5;
  s1 [label="Redistribute\n1024 / 1024 cycles\n{}",color=black];
  v5 -> s1;
  s1 -> v6;
  s2 [label="CastAndPack\n522 / 2088 cycles\n{MissingKernel(\"F16 RowMajor -> F8F143 {\nscale_exponent: -4 } BlockMajor(TransposedMatrix {\nrow_block: 32, column_block: 16 })\")}",color=red];
  v6 -> s2;
  s2 -> v7;
  s3 [label="Redistribute\n1020 / 1020 cycles\n{}",color=black];
  v3 -> s3;
  s3 -> v8;
  s4 [label="CastAndPack\n426 / 1704 cycles\n{MissingKernel(\"F16 RowMajor -> F8F143 {\nscale_exponent: -4 } Amp(TransposedLeft)\")}",color=red];
  v8 -> s4;
  s4 -> v9;
  s5 [label="Gemm { options: GemmOptions { transpose_left: false,\ntranspose_right: false }, multiply: F8F143 {\nscale_exponent: -4 }, accumulate: F16 }\n2902 / 2902 cycles\n{}",color=black];
  v7 -> s5;
  v9 -> s5;
  s5 -> v10;
  s6 [label="Redistribute\n1020 / 1020 cycles\n{}",color=black];
  v4 -> s6;
  s6 -> v11;
  s7 [label="CastAndPack\n426 / 1704 cycles\n{MissingKernel(\"F16 RowMajor -> F8F143 {\nscale_exponent: -4 } Amp(TransposedLeft)\")}",color=red];
  v11 -> s7;
  s7 -> v12;
  s8 [label="Gemm { options: GemmOptions { transpose_left: false,\ntranspose_right: false }, multiply: F8F143 {\nscale_exponent: -4 }, accumulate: F16 }\n2902 / 2902 cycles\n{}",color=black];
  v7 -> s8;
  v12 -> s8;
  s8 -> v13;
  s9 [label="Pack\n346 / 346 cycles\n{}",color=black];
  v10 -> s9;
  s9 -> v14;
  s10 [label="Redistribute\n952 / 952 cycles\n{}",color=black];
  v14 -> s10;
  s10 -> v15;
  s11 [label="Pack\n346 / 346 cycles\n{}",color=black];
  v13 -> s11;
  s11 -> v16;
  s12 [label="Redistribute\n952 / 952 cycles\n{}",color=black];
  v16 -> s12;
  s12 -> v17;
  v15 [penwidth=3];
  v17 [penwidth=3];
}