Add Support for intfloat/multilingual-e5-large-instruct :Fixes #140

[Ya-shh]

Description:

This pull request addresses issue #140 by integrating the intfloat/multilingual-e5-large-instruct model from Hugging Face into the repo. This update extends capabilities, enabling it to support more diverse language embeddings and accommodate various numerical data types seamlessly.

Changes:

•Model Support: The intfloat/multilingual-e5-large-instruct model is now supported, expanding our project's multilingual
processing capabilities.

•Documentation Update: Updated the Supported_Models.ipynb notebook to include documentation and usage examples for the new model.

•Configuration Update: Modified e5_onnx_embedding.py to include configuration settings specific to the intfloat/multilingual-e5-large-instruct model, ensuring optimal performance.

•Testing: Enhanced test_text_onnx_embeddings.py with new test cases designed to validate the output embeddings of the intfloat/multilingual-e5-large-instruct model against predefined documents.

•CANONICAL_VECTOR values : Here is a reference Colab Notebook

Onnx file :

HuggingFaceHub:
https://huggingface.co/yashvardhan7/multilingual-e5-large-instruct/tree/main

Test:

All test cases have successfully passed with the inclusion of the latest intfloat/multilingual-e5-large-instruct model.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[Ya-shh]

@NirantK I have added Huggingface url which contains onnx model file, also both the tests are passed.

[Ya-shh]

@NirantK could you please review the changes?

[Anush008]

Hello @Ya-shh. You've gated your model on HF, which is why a3a3e06 failed. The download should be good after you remove it.

cc58940 won't be necessary.

[Ya-shh]

Hello @Ya-shh. You've gated your model on HF, which is why a3a3e06 failed. The download should be good after you remove it.

cc58940 won't be necessary.

Thanks a lot ! @Anush008 . I have disabled gated user access, you could try it again.

[Ya-shh]

Hey @Anush008 , now it won't fail . You could run it !

[Ya-shh]

@NirantK all checks have passed !

[Ya-shh]

@Anush008 I have committed the changes .Now it's ready !

[Anush008]

Thank you.

[Anush008]

The Torch canonical values you've shared in the notebook seem to be different from the ones in the test.

[Ya-shh]

Hi @Anush008, In reference to the issue discussed in #190, I have identified the error in this (PR) as being related to the atol values. Initially, I incorrectly modified the canonical vector values; however, I've now made the necessary corrections to the atol values. Additionally, I am addressing the token_type_ids issue mentioned in #190 (currently marked as draft). All other concerns raised regarding this PR have been resolved.

[Ya-shh]

@NirantK, All the issues with this PR have been resolved.

[NirantK]

No, we don't want to use such large atol values. 1e-3 is already quite large — please find an ONNX model which clears the tests.

[Ya-shh]

Yes , I totally agree with you. But even with the e5-small Tests failed at all close

[Anush008]

You can maybe remove the other models when trying the tests locally.

[Ya-shh]

@NirantK I have tried every possible way to quantise this model .Optimized it with different levels graph optimizations . And the result remain the same for O4, O3 and O2 with an atol=1e-3

E AssertionError: intfloat/multilingual-e5-large-instruct
E assert False
E + where False = <function allclose at 0x101d18270>(array([ 0.01044647, 0.02897521, 0.00060697, -0.04193534, 0.02606635],\n dtype=float32), array([ 0.01020065, 0.02367218, 0.00117696, -0.04327099, 0.0288757 ]), atol=0.001)
E + where <function allclose at 0x101d18270> = np.allclose

I have even quantized this model further using Static quantization but still no change in the onnx weights .It only passes for atol=1e-2. As you said 1e-3 is already quite large value. So we really can’t add support for this model.

[Ya-shh]

Here is the Onnx model

CANONICAL_VECTOR values :Colab Notebook

[Ya-shh]

I have tried all of these but still these fails at atol=1e-3. But passes at atol=1e-2 while testing them locally.

Colab notebooks

"O2" level graph optimizations

"O3" level graph optimizations

"O4" level graph optimizations

Onnx/export space

Onnx model using Onnx/export space on HF

[Ya-shh]

@NirantK, locally I have tested both Xenova/multilingual-e5-small & intfloat/multilingual-e5-small onnx models but they also fails at allclose : #123

E AssertionError: intfloat/multilingual-e5-small
E assert False
E + where False = <function allclose at 0x110ad0170>(array([ 0.04931236, 0.02415175, -0.0384715 , -0.08884481, 0.08710264],\n dtype=float32), array([ 0.03131689, 0.03093922, -0.03511665, -0.06727391, 0.08508426]), atol=0.001)
E + where <function allclose at 0x110ad0170> = np.allclose

But unlike e5-large-instruct , e5-small also fails at atol=1e-2

E AssertionError: intfloat/multilingual-e5-small
E assert False
E + where False = <function allclose at 0x10456d470>(array([ 0.04931236, 0.02415175, -0.0384715 , -0.08884481, 0.08710264],\n dtype=float32), array([ 0.03131689, 0.03093922, -0.03511665, -0.06727391, 0.08508426]), atol=0.01)
E + where <function allclose at 0x10456d470> = np.allclose

These multilingual models really require more investigation.

[Ya-shh]

@NirantK, I am closing this PR for now. Whenever you have time to review, kindly take the time to go through all the attached notebooks. Your thorough review will greatly assist me in addressing any issues with these models. Thank you for your valuable suggestions!