return nullptr;
  }
  auto mins = tensor.quantization->min;
  auto maxs = tensor.quantization->max;
  if (mins.size() != maxs.size() || mins.empty()) return nullptr;

  llvm::SmallVector<llvm::APFloat, 4> min_maxs;
  min_maxs.reserve(mins.size() * 2);
  for (int i = 0, end = mins.size(); i < end; ++i) {
    llvm::APFloat min(mins[i]);
    llvm::APFloat max(maxs[i]);
    min_maxs.push_back(min);

          "interval": "1m",
          "legendFormat": "{{server}}",
          "refId": "A"
        }
      ],
      "title": "Backlog (last 5 mins)",
      "type": "timeseries"
    },
    {
      "datasource": {
        "type": "prometheus",
        "uid": "${DS_PROMETHEUS}"
      },
      "fieldConfig": {
        "defaults": {

		{name: "FMIND", argLength: 2, reg: fp21, asm: "FMIND"},                                // min(arg0, arg1)
		{name: "FMINS", argLength: 2, reg: fp21, asm: "FMINS"},                                // min(arg0, arg1)
		{name: "FMAXD", argLength: 2, reg: fp21, asm: "FMAXD"},                                // max(arg0, arg1)

// Note that this trick depends on the special property that (NaN OR x) produces a NaN (although
// it might not produce the same NaN as the input).
(Min(64|32)F <t> x y) => (POR (MINS(D|S) <t> (MINS(D|S) <t> x y) x) (MINS(D|S) <t> x y))
// Floating-point max is even trickier. Punt to using min instead.
// max(x,y) == -min(-x,-y)
(Max(64|32)F <t> x y) => (Neg(64|32)F <t> (Min(64|32)F <t> (Neg(64|32)F <t> x) (Neg(64|32)F <t> y)))

             'compilation to reduce the compilation time?'),
      'Eigen strong inline overridden.', 'Not overriding eigen strong inline, '
      'some compilations could take more than 20 mins.'):
    # Due to a known MSVC compiler issue
    # https://github.com/tensorflow/tensorflow/issues/10521
    # Overriding eigen strong inline speeds up the compiling of