LSTM model failed on macos

[Lizhen0909]

I ran a simple lstm code as below on macos with torch backend. It showed error:

RuntimeError: Unable to automatically build the model. Please build it yourself before calling fit/evaluate/predict. A model is 'built' when its variables have been created and its `self.built` attribute is True. Usually, calling the model on a batch of data is the right way to build it.
Exception encountered:
'Exception encountered when calling LSTM.call().

'LSTMCell' object has no attribute 'recurrent_kernel'

Arguments received by LSTM.call():
  • sequences=torch.Tensor(shape=torch.Size([32, 10, 8]), dtype=float32)
  • initial_state=None
  • mask=None
  • training=False'

environment:

Python: 3.11.4
Torch version: 2.8.0
Keras version: 3.11.3

I tried also on macos with tensorflow backend which had no problem. Also tried on linux with torch backend which was also good.
Is this a bug or my env problem?

Sample code:

import numpy as np
import os 
os.environ['KERAS_BACKEND'] = 'torch'
import keras
from keras import layers
import torch
print("Torch version:", torch.__version__)
print("Keras backend:", keras.backend.backend())
print("Keras version:", keras.__version__)

# Generate dummy sequential data
x_train = np.random.random((1000, 10, 8))  # 1000 samples, 10 timesteps, 8 features
y_train = np.random.randint(2, size=(1000, 1))  # Binary targets

# Build" LSTM model
model = keras.Sequential([
    layers.LSTM(32, input_shape=(10, 8)),
    layers.Dense(1, activation='sigmoid')
])

model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])

# Train model
model.fit(x_train, y_train, epochs=5, batch_size=32)

# Evaluate model
x_test = np.random.random((100, 10, 8))
y_test = np.random.randint(2, size=(100, 1))
loss, acc = model.evaluate(x_test, y_test)
print(f"Test loss: {loss:.4f}, Test accuracy: {acc:.4f}")

[SanchitKS12]

I would like to make a suggestion maybe there is actually a issue with your env(causing partial compilation error) or maybe backend mismatches ,because if you look closely and also as you have mentioned it works correctly with macOS + TensorFlow backend and with Linux + Torch backend which clearly justifies the fact that their is no code issue what so ever

'LSTMCell' object has no attribute 'recurrent_kernel'

The error statement itself points towards the same and since it works on Linux with the same versions, it’s not a code bug in Keras itself — more likely a PyTorch build or ABI mismatch on macOS.
What you can do is test a minimal non-LSTM model (e.g., Dense only). If that runs fine, it confirms the problem is isolated to the Torch LSTMCell implementation on macOS.
Hope it helps and please let me know if i am wrong somewhere.