// NoFusing-LABEL: FuseMulWithFullyConnectedNoBias
  // NoFusing-DAG: %[[MWEIGHTS:.*]] = arith.constant dense<2.000000e+00> : tensor<512xf32>
  // NoFusing-DAG: %[[WEIGHTS:.*]] = arith.constant dense<3.000000e+00> : tensor<1024x512xf32>
  // NoFusing-DAG: %[[BIAS:.*]] = "tfl.no_value"() <{value}> : () -> none

          &q_builder, input_model, quantized_type, use_updated_hybrid_scheme,
          ::tflite::optimize::QuantizerType::OLD_QUANTIZER) != kTfLiteOk) {
    return absl::InvalidArgumentError(
        "Quantize weights transformation failed.");
  }
  const uint8_t* q_buffer = q_builder.GetBufferPointer();
  *result =
      std::string(reinterpret_cast<const char*>(q_buffer), q_builder.GetSize());

  return absl::OkStatus();
}

    Isso significa que, embora os clientes da API só possam enviar strings como chaves, desde que essas strings contenham inteiros puros, o Pydantic irá convertê-los e validá-los.

    E o `dict` que você recebe como `weights` terá, na verdade, chaves `int` e valores` float`.

## Recapitulação

Com **FastAPI** você tem a flexibilidade máxima fornecida pelos modelos Pydantic, enquanto seu código é mantido simples, curto e elegante.

inline constexpr std::array<int64_t, 4> kNchwToNhwcPermutation = {0, 2, 3, 1};

// Permutation from the OIHW (== (output features, input features, height,
// width)) tensor format to HWIO. This is commonly used to transpose convolution
// weights represented as OIHW format to HWIO, which is more desirable for
// certain downstream optimization passes (e.g. XLA).

                                    enable_per_channel_quantization_));

  // Apply activation-weight quantization.
  if (quantization_method_ ==
      tensorflow::quantization::QuantizationMethod::METHOD_STATIC_RANGE_INT8) {
    // For XLA case, weight quantization will be applied for the remaining f32
    // weights even in SRQ.
    pm.addNestedPass<func::FuncOp>(

  SPARSE = 1,
  DENSE = 2,
}

table LSHProjectionOptions {
  type: LSHProjectionType;
}

table SVDFOptions {
  rank:int;
  fused_activation_function:ActivationFunctionType;
  // For weights-only quantization, use asymmetric quantization for non
  // constant inputs at evaluation time.
  asymmetric_quantize_inputs:bool;
}

// An implementation of TensorFlow RNNCell.
table RNNOptions {

}

func (e *LocalityEndpoints) refreshWeight() {
	var weight *wrapperspb.UInt32Value
	if len(e.llbEndpoints.LbEndpoints) == 0 {
		weight = nil
	} else {
		weight = &wrapperspb.UInt32Value{}
		for _, lbEp := range e.llbEndpoints.LbEndpoints {
			weight.Value += lbEp.GetLoadBalancingWeight().Value
		}
	}
	e.llbEndpoints.LoadBalancingWeight = weight
}

func (e *LocalityEndpoints) AssertInvarianceInTest() {

              "Non-constant weights are not supported at the moment,"
              " except matmul and einsum.");
        } else if (!quant_options_.enable_two_input_tensors() &&
                   !is_unitwise_quantization_enabled) {
          return absl::InternalError(
              "Quantization is disabled for this op due to the non-constant "
              "weight. You can enable it by setting `enable_two_input_tensors` "

// RUN: tf-opt %s -tfl-prepare-quantize-dynamic-range="min-elements-for-weights=4000 enable-custom-op-quantization=CustomTestOp=1-3,CustomTestOp3=3" | FileCheck --check-prefix=MinElement %s
// RUN: tf-opt %s -tfl-prepare-quantize-dynamic-range="min-elements-for-weights=19" | FileCheck --check-prefix=LSTMOpQuantized %s
// RUN: tf-opt %s -tfl-prepare-quantize-dynamic-range="min-elements-for-weights=21" | FileCheck --check-prefix=LSTMOpNotQuantized %s

```

!!! tip "팁"
    JSON은 오직 `str`형 키만 지원한다는 것을 염두에 두세요.

    하지만 Pydantic은 자동 데이터 변환이 있습니다.

    즉, API 클라이언트가 문자열을 키로 보내더라도 해당 문자열이 순수한 정수를 포함하는한 Pydantic은 이를 변환하고 검증합니다.

    그러므로 `weights`로 받은 `dict`는 실제로 `int` 키와 `float` 값을 가집니다.

## 요약

**FastAPI**를 사용하면 Pydantic 모델이 제공하는 최대 유연성을 확보하면서 코드를 간단하고 짧게, 그리고 우아하게 유지할 수 있습니다.

물론 아래의 이점도 있습니다:

* 편집기 지원 (자동완성이 어디서나!)
* 데이터 변환 (일명 파싱/직렬화)