EXPECT_EQ(*result_value, 1.0);
  outputs[0]->Unref();
  TF_DeleteTensor(result_tensor);
  result_tensor = nullptr;
}

INSTANTIATE_TEST_SUITE_P(
    CustomGradientTest, CustomGradientTest,
    ::testing::Combine(::testing::Values("graphdef", "mlir"),
                       /*tfrt*/ ::testing::Values(false),
                       /*executing_eagerly*/ ::testing::Values(true, false)));

}  // namespace
}  // namespace internal

// argument, even if it has the default value (this means Types<int>
// will appear as Types<int, None, None, ..., None> in the compiler
// errors).
//
// Our solution is to combine the best part of the two approaches: a
// user would write Types<T1, ..., TN>, and Google Test will translate
// that to TypesN<T1, ..., TN> internally to make error messages

void PopulateLegalizeTfPatterns(MLIRContext* context,
                                RewritePatternSet* patterns);

// Populates TF to MHLO legalization for some of the quantization ops.
//
// TODO(hinsu): Remove this once we combine quantized and non quantized op
// legalization in the ODML conversion pipeline.
void PopulateLegalizeTfQuantizationPatterns(MLIRContext* context,
                                            RewritePatternSet* patterns);

// Copyright 2015 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.

package ssa

// combine copyelim and phielim into a single pass.
// copyelim removes all uses of OpCopy values from f.
// A subsequent deadcode pass is needed to actually remove the copies.
func copyelim(f *Func) {
	phielim(f)

                throw UncheckedException.throwAsUncheckedException(t);
            } else {
                // Created a result which may contain failures. Combine this failure with any failures that happen to be packaged in the result

                throw UncheckedException.throwAsUncheckedException(t);
            } else {
                // Cleanup has failed, combine the cleanup failure with other failures that may be packed in the result

  ASSERT_FALSE(shape.unknown_rank());

  ASSERT_EQ(2, shape.dim_size(0));
  ASSERT_EQ(-1, shape.dim_size(1));
}

INSTANTIATE_TEST_SUITE_P(
    UnifiedCppAPI, UnifiedAPI,
    ::testing::Combine(::testing::Values("graphdef", "mlir"),
                       /*tfrt*/ ::testing::Values(false),
                       /*use_function*/ ::testing::Values(true, false)));
}  // namespace

</blockquote>

## Investigation

By using all the previous alternatives I had the chance to learn from all of them, take ideas, and combine them in the best way I could find for myself and the teams of developers I have worked with.

For example, it was clear that ideally it should be based on standard Python type hints.

```

Tout sera combiné et inclus dans votre OpenAPI, et affiché dans la documentation de l'API :

<img src="/img/tutorial/additional-responses/image01.png">

## Combinez les réponses prédéfinies et les réponses personnalisées

      return errors::Internal("Unknown dimension size in constant tensor ",
                              n->name());
    }
    total_elements *= dim.size();
  }

  // TODO(sanjoy): It may make sense to combine this threshold with XLA's "large
  // constant" threshold, if there is one.
  const int kSmallTensorThreshold = 16;
  return total_elements < kSmallTensorThreshold;
}