Skip to main content
Question

Vertica 25.4.1 ERROR 9450: Unsupported tensorflow type: 0 when importing Universal Sentence Encoder (Multilingual)

  • Feb 10, 2026 12:43 AM UTC
  • 3 replies
  • 26 views

steven_chang_tradevan_com_tw_

  

Hi everyone,

I am struggling to import the Google Universal Sentence Encoder (Multilingual) model into Vertica 25.4.1 using the TensorFlow integration.

My Environment:

  • OS: Rocky Linux 9

  • Vertica Version: 25.4.1-x

  • TensorFlow: 2.15.0

  • TensorFlow Text: 2.15.0 (required for the multilingual USE ops)

  • Python: 3.10 (for model freezing)

The Issue: I have successfully loaded the USE model in Python, registered the SentencepieceOp via tensorflow_text, and exported it as a frozen .pb file with a corresponding tf_model_desc.json.

However, when I run the following SQL command in Vertica:

SELECT IMPORT_MODELS('/home/dbadmin/frozen_dir/' USING PARAMETERS category='TENSORFLOW');

I consistently get this error: ERROR 9450: Failed to import model(s): Failed to import single model: Error in importModelFiles: Error calling setup() in User Function import_model_files at [src/Common/TFCommon.cpp:51], error code: 0, message: Unsupported tensorflow type: 0

My tf_model_desc.json:

{
    "frozen_graph": "use_model.pb",
    "input_desc": [{
        "name": "input_text",
        "dims": [-1],
        "data_type": "string"
    }],
    "output_desc": [{
        "name": "Identity",
        "dims": [-1, 512],
        "data_type": "float"
    }]
}

What I've tried:

  1. Verified that the tensor name input_text exists in the frozen graph.

  2. Attempted changing data_type in the JSON to "STRING", "string", and tried changing dims from [-1] to [1].

  3. Confirmed that tensorflow_text ops are required for this model.

The error type: 0 seems to map to DT_INVALID in the TensorFlow source. Does Vertica's TensorFlow integration support DT_STRING tensors for the Multilingual USE model? Or is there a specific way to define string inputs in the descriptor file for this version of Vertica?

Any guidance or workarounds would be greatly appreciated!

3 replies

SruthiA
Forum|alt.badge.img+1
  • Participating Frequently
  • Feb 10, 2026 1:54 PM UTC

No it is not supported.   Please check the below link on supported datatype. Please raise support case and share sample example, we can create new feature request.

docs.vertica.com/.../


steven_chang_tradevan_com_tw_

Hi ,

  Thank you for your help. 

Modern NLP tasks require technologies that exceed Vertica's native capabilities:

  • Native Limitations: Vertica's native TensorFlow integration is restricted to TensorFlow 1.15. Analysis of symbols in libMachineLearning.so confirms this; while symbols containing the B5cxx11 suffix indicate that Vertica is compiled with a modern C++11 ABI, the presence of legacy TF_INT32 and TF_FLOAT symbols confirms that the embedded engine relies on the original TensorFlow 1.x C API. This restricts native execution to old "Frozen Graph" (.pb) models.

  • PMML Limitations: PMML support in Vertica is limited to traditional statistical models (e.g., Regression, Tree-based models) and cannot describe the complex Self-Attention mechanisms required by Transformers.

Consequently, Python UDX with ONNX Runtime is the only viable path for deploying modern state-of-the-art models.

Regards,

Steven


SruthiA
Forum|alt.badge.img+1
  • Participating Frequently
  • Feb 26, 2026 7:33 PM UTC

thank you for sharing detailed info.. Please raise support case and share above info. we can create new feature request.