[Bugfix] Point OnnxGz files to model.onnx but download and extract model.onnx.tar.gz

[rahul-tuli]

Big LLMS are generally saved with external data; this external data along with model.onnx live inside model.onnx.tar.gz tarball on sparsezoo; recently we added support to download the tarfile and emulate a normal onnx model file using a wrapper class OnnxGz; when adding this support a bug was introduced where model.onnx was present twice in the model path; leading to errors while creating ort sessions as

TEST PLAN:

• Manual inspection using attached script
• Add automated tests to capture conditions where extraction was failing
• Test with deepsparse main

Test Script:

from sparsezoo import Model
from pathlib import Path


def local_test(stub: str):
    model = Model(stub)
    onnx_model = model.onnx_model
    model_path = onnx_model.path
    onnx_model_path = Path(model_path)
    assert onnx_model_path.exists(), "Path to onnx model does not exist"
    assert (count := str(onnx_model_path).count("model.onnx")) == 1, f"Path to onnx model contains model.onnx {count} times"

    # create an ORT session with the model
    import onnxruntime as ort
    session = ort.InferenceSession(str(onnx_model_path), providers=["CPUExecutionProvider"])
    assert session is not None, "ORT session is None"

if __name__ == '__main__':
    # stub = "zoo:nlg/text_generation/opt-13b/pytorch/huggingface/opt_pretrain/pruned50_quantW8A8-none"
    stub = "zoo:cv/classification/mobilenet_v1-1.0/pytorch/sparseml/imagenet/pruned_quant-moderate"
    local_test(stub=stub)

Before this PR: The onnx_model_path had an extra model.onnx in it's path leading to errors like:
https://github.com/neuralmagic/deepsparse/actions/runs/5869604261/job/15914935604?pr=1150#step:7:138

After this PR the test script runs successfully, and the base test errors on deepsparse side are green

---

• To see the specific tasks where the Asana app for GitHub is being used, see below:
  • https://app.asana.com/0/0/1205276886236935

[Satrat]

I'm pretty sure the file we want to unzip is model.tar.gz and not model.onnx.tar.gz if the purpose of this PR is to extract the model.data file.

The 2B codegen stub shows model.tar.gz as 6.4GB and model.onnx.tar.gz as only 594KB.

In general, I think it would be a good idea to upload a small model to sparsezoo that has its data stored externally so we can catch the issue in the test suite without having to run giant models

[rahul-tuli]

I'm pretty sure the file we want to unzip is model.tar.gz and not model.onnx.tar.gz if the purpose of this PR is to extract the model.data file.

The 2B codegen stub shows model.tar.gz as 6.4GB and model.onnx.tar.gz as only 594KB.

In general, I think it would be a good idea to upload a small model to sparsezoo that has its data stored externally so we can catch the issue in the test suite without having to run giant models

making a comment here that the linked model needs to be updated on sparsezoo; and the onnx model model.onnx along with external_data if any should be inside model.onnx.tar.gz

[dsikka]

Two comments. otherwise LGTM