keras-team/keras · error · AttributeError

You must build the layer before accessing `kernel`.

Error message

You must build the layer before accessing `kernel`.

What it means

EinsumDense.kernel is a property over the weight variable, which is only created when build() runs (weights are lazy, derived from input shape and equation resolution). Accessing .kernel before build raises AttributeError since no variable exists.

Source

Thrown at keras/src/layers/core/einsum_dense.py:242

                shape=tuple(bias_shape),
                initializer=self.bias_initializer,
                regularizer=self.bias_regularizer,
                constraint=self.bias_constraint,
                dtype=self.dtype,
                trainable=True,
            )
        else:
            self.bias = None
        self.built = True
        if self.lora_rank:
            self.enable_lora(self.lora_rank, lora_alpha=self.lora_alpha)

    @property
    def kernel(self):
        from keras.src.quantizers import gptq_core

        if not self.built:
            raise AttributeError(
                "You must build the layer before accessing `kernel`."
            )

        mode = self.quantization_mode
        is_gptq = mode == "gptq"
        is_awq = mode == "awq"
        is_int4 = mode == "int4"
        gptq_calibrated = bool(getattr(self, "is_gptq_calibrated", False))
        awq_calibrated = bool(getattr(self, "is_awq_calibrated", False))
        gptq_bits = (
            gptq_core.get_weight_bits_for_layer(self, None) if is_gptq else None
        )

        # Decide the source tensor first (packed vs already-quantized vs plain
        # kernel)
        if is_gptq and gptq_calibrated and gptq_bits not in (2, 4):
            # calibrated GPTQ, not a packed bit-width, no unpacking needed
            kernel = self.quantized_kernel

View on GitHub (pinned to 7a34a03db6)

Solutions

1. Build the layer first: layer.build(input_shape) or run a forward pass, then access .kernel.
2. Use compute_output_shape / equation shape utilities instead of reading the kernel for shape logic.
3. For whole models, call model.build(input_shape) before inspecting layer weights.

Example fix

# before
layer = keras.layers.EinsumDense('ab,bc->ac', output_shape=(None, 64))
print(layer.kernel.shape)  # AttributeError

# after
layer = keras.layers.EinsumDense('ab,bc->ac', output_shape=(None, 64))
layer.build((None, 32))
print(layer.kernel.shape)

if not layer.built:
    layer.build(input_shape)
k = layer.kernel

def safe_kernel(layer):
    return layer.kernel if layer.built else None

When it happens

Trigger: Reading layer.kernel (or .kernel.shape) on an EinsumDense that hasn't seen an input shape — before layer(input) or layer.build(input_shape).

Common situations: Weight-initialization or inspection code right after construction; custom loading logic reading kernel before build; functional models probed before inputs resolve.

Related errors

• You must build the layer before accessing `kernel`.
• Cannot enable lora on a layer that isn't yet built.
• Lora is incompatible with kernel constraints. In order to en
• Cannot enable lora on a layer that isn't yet built.
• lora is already enabled. This can only be done once per laye

AI-assisted analysis of keras-team/keras@7a34a03db6 (2026-08-25). Data as JSON: /api/errors/7fb0bf584725fb88. Report an issue: GitHub.