Fix mixtral ONNX Exporter Issue.

[AdamLouly]

This PR addresses issue #29857

by removing the list conversion of top_x and idx in the Mixtral model code, which currently obstructs the ONNX exporter's functionality.

By removing this conversion, the Mixtral model can be exported successfully using the ONNX exporter, resolving the compatibility issue.

[AdamLouly]

@thiagocrepaldi FYI.

[thiagocrepaldi]

to expedite the code review, use git blame and find previous users who actually approved prs for changes in this file and add them to the code review.

Otherwise, the PR can be waiting for triage a long time

@ArthurZucker FYI

[ArthurZucker]

Thanks a lot!
Would love to merge this, we had to go with this because indexing is faster with list.

would you mind just running some benchmarks in generation?
Really sorry we don't have them yet in transformers but coming soon! (FYI @ydshieh )

[AdamLouly]

Hey @ArthurZucker any updates on this?

[ydshieh]

@AdamLouly

I think what @ArthurZucker 's comment is about

would you mind just running some benchmarks in generation?

on your own side to see if the PR and main have difference in the speed of generation for this model

[ArthurZucker]

On point @ydshieh . This is very core, and we added this change for a reason, we need a good reason to update, and make sure we are not affecting performances

[AdamLouly]

@ArthurZucker, I actually found it faster when we use the changes in the PR.

Environment specs:
V100
Mixtral-8x7B-v0.1
2 Hidden layers
Loaded using torch_dtype=torch.float16

With the PR changes:
{'train_runtime': 182.1736, 'train_samples_per_second': 8.783, 'train_steps_per_second': 0.274, 'train_loss': 1.3467385864257813, 'epoch': 5.71}

Without PR changes:
{'train_runtime': 200.6888, 'train_samples_per_second': 7.973, 'train_steps_per_second': 0.249, 'train_loss': 1.3390922546386719, 'epoch': 5.71}

Thank you

[ArthurZucker]

Alright. Let's also update the qwen models as well then! (ctrl + f for this operation)

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[AdamLouly]

I could not find anything like that on qwen models, are you sure it uses the same thing?

[ArthurZucker]

This here! :

transformers/src/transformers/models/qwen2_moe/modeling_qwen2_moe.py

Line 847 in 8e51901

top_x_list = top_x.tolist()

qwen2 sorry

[ArthurZucker]

Thanks