Bug report

keras.layers.SeparableConv1D computes a correct output shape for eager tensors, but the symbolic output shape from keras.Input() is wrong when using padding='valid' and strides > 1.

Reproduction

import keras
import tensorflow as tf
import numpy as np

x_eager = tf.constant(np.random.rand(4, 10, 12), dtype=tf.float32)
x_sym = keras.Input(shape=(10, 12))

layer = keras.layers.SeparableConv1D(64, 11, strides=5, padding='valid', activation='relu')

y_eager = layer(x_eager)
y_sym = layer(x_sym)

print(y_eager.shape)
print(y_sym.shape)

Expected result

Both outputs should report the same spatial shape.

Actual result

• Eager output shape: (4, 10, 64)
• Symbolic output shape: (None, 0, 64)

This indicates incorrect static shape inference for SeparableConv1D.