`Conv1D` silently returns zero-dimension output in symbolic mode instead of raising an error for invalid parameters.

[AGFACBNNR]

Related Keras Version
keras==3.13.1, keras==3.10.0
The Conv1D layer shows inconsistent behavior between eager and symbolic execution. When given parameters that are invalid for the input shape (e.g., a dilation_rate that makes the effective kernel larger than the input), it correctly raises an error in eager mode.

However, in symbolic mode, it fails silently and returns a KerasTensor with a zero-dimension shape (e.g., (None, 0, 1)). This can mask configuration errors during model construction, leading to downstream issues that are difficult to debug. The symbolic path should perform the same validation and fail fast.

This issue is observed across different backends (TensorFlow==2.20.0, PyTorch==2.10.0+cu128).

Code to Reproduce

import keras

print(f"Keras version: {keras.__version__}")

# Eager execution
try:
    eager_input = keras.random.normal(shape=(2, 10, 3))
    layer = keras.layers.Conv1D(
        filters=1, 
        kernel_size=2, 
        strides=1, 
        padding='valid',
        dilation_rate = 10,
        activation="relu"
    )
    eager_output = layer(eager_input)
    print("Eager output shape:", eager_output.shape) 
except Exception as e:
    print("Error during eager execution:", e)

# Symbolic execution
symbolic_input = keras.Input(shape=(10, 3)) 
symbolic_output = layer(symbolic_input)
print("Symbolic output shape:", symbolic_output.shape)

Actual Behavior

Eager execution fails with a descriptive error, but symbolic execution proceeds and produces a tensor with a zero dimension.

Keras version: 3.13.1
Error during eager execution: "The convolution operation resulted in an empty output. Output shape: (2, 0, 1)..."
Symbolic output shape: (None, 0, 1)

(Note: The exact error message varies by backend, but the eager failure is consistent.)

Expected Behavior

Symbolic execution should also raise an error during graph construction, mirroring the eager behavior. This would provide immediate feedback that the layer configuration is invalid.

Keras version: 3.13.1
Error during eager execution: "The convolution operation resulted in an empty output..."
# Followed by a similar error during the symbolic call:
# layer(symbolic_input)
# -> raises ValueError: "The convolution parameters are invalid for the given input shape..."

[AGFACBNNR]

It is worth noting that for the TensorFlow backend, the eager execution error described above is specific to the GPU implementation.

When forcing CPU execution with os.environ["CUDA_VISIBLE_DEVICES"] = "-1", the eager path does not raise an error. Instead, it silently returns a tensor with an incorrect shape of (2, 10, 3).

I will file a separate issue in the tensorflow/tensorflow repository to report this inconsistent behavior between their CPU and GPU operations. The primary issue for Keras remains the discrepancy between symbolic and eager execution paths.