New out-of-the-box error as of May 2025 (tf version 2.19.0)

[jasonnett80]

I've been coming back to this example for testing for a few years now and there is a new error that this code now produces out-of-the-box that it did not as of a few months ago (winter 2025). This is in reference to the english-to-spanish machine translation example.

https://keras.io/examples/nlp/neural_machine_translation_with_transformer/

inputs["encoder_inputs"].shape: (64, 20)
inputs["decoder_inputs"].shape: (64, 20)
targets.shape: (64, 20)
2025-05-21 17:05:53.777692: I tensorflow/core/framework/local_rendezvous.cc:407] Local rendezvous is aborting with status: OUT_OF_RANGE: End of sequence
Traceback (most recent call last):
  File "/home/dev_account/project/transformer_keras/transformer_keras_original.py", line 417, in <module>
    decoder = keras.Model([decoder_inputs, encoded_seq_inputs], decoder_outputs)
  File "/usr/local/lib/python3.10/dist-packages/keras/src/utils/tracking.py", line 26, in wrapper
    return fn(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/keras/src/models/functional.py", line 136, in __init__
    Function.__init__(self, inputs, outputs, name=name)
  File "/usr/local/lib/python3.10/dist-packages/keras/src/ops/function.py", line 89, in __init__
    raise ValueError("`inputs` not connected to `outputs`")
ValueError: `inputs` not connected to `outputs`

[sachinprasadhs]

Hi, What is the TensorFlow and Keras version you are using, also are you using the keras-hub library?

[jasonnett80]

@sachinprasadhs

Tensorflow 2.19.0 and keras 3.10. Negative on keras-hub.

[jasonnett80]

I think I found that the solution was to replace this

decoder = keras.Model([decoder_inputs, encoded_seq_inputs], decoder_outputs)

with

decoder = keras.Model(decoder_inputs, decoder_outputs)

but that begs the question: what functionality does encoded_seq_inputs provide and what are the consequences of removing it? Executing as such, translation results appear to be about equivalent.

[github-actions]

This issue is stale because it has been open for 14 days with no activity. It will be closed if no further activity occurs. Thank you.

[alrahimi]

I think I found that the solution was to replace this

decoder = keras.Model([decoder_inputs, encoded_seq_inputs], decoder_outputs)

with

decoder = keras.Model(decoder_inputs, decoder_outputs)

but that begs the question: what functionality does encoded_seq_inputs provide and what are the consequences of removing it? Executing as such, translation results appear to be about equivalent.

[alrahimi]

I found the same solution/work around. I didn't know about #2017 until I filed a new issue (Issue #2175) and got response:
On Mon, Sep 8, 2025 at 2:53 AM Sonali Kumari <notifications@github.com> wrote:

sonali-kumari1
left a comment
(keras-team/keras-io#2175)
Hi @alrahimi -
This issue appears to be a duplicate of issue #2107. Please feel free to follow the linked issue for updates. Thanks!

[alrahimi]

Here is an updated version of neural machine translation with transformer It builds. Trained successfully for one Epoch!
But translations are relevant!
All the outputs are included

neural_machine_translation_with_transformer_arahimi_R01_01.ipynb

[alrahimi]

I think I found that the solution was to replace this

decoder = keras.Model([decoder_inputs, encoded_seq_inputs], decoder_outputs)

with

decoder = keras.Model(decoder_inputs, decoder_outputs)

but that begs the question: what functionality does encoded_seq_inputs provide and what are the consequences of removing it? Executing as such, translation results appear to be about equivalent.

The decoder itself is never used! (the TransformerDecoder is the real decoder). I removed it in my version (see the attached Collab notebook in my response), and everything works