Fix Qwen3

[kozistr]

What does this PR do?

Fixes #642

Changes

• left padding (w/ pad token id)
• causal attention mask
• fix last token pooling for the batch case

• Currently, batch inference works fine, but single-query inference has a problem.
• (updated) I've further investigated this issue and found that it's due to the attention logic.
  • self.config._attn_implementation of Qwen3 Embedding config is sdpa by default, and if we change this value to eager, which is naive impl of self-attention, we could check the identical output with this PR as expected.
  • but, looks like sdpa performs something different, need to investigate more on this

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@Narsil @alvarobartt

[alvarobartt]

Thanks for the PR @kozistr and apologies I missed those points! LGTM

[kozistr]

@alvarobartt @Narsil thanks for the quick check!

I've further investigated this issue (== single query inference seems weird) and found that it's due to the attention logic. self.config._attn_implementation of Qwen3 Embedding config is sdpa by default, and if we change this value to eager, which is naive implementation of self-attention, we could check the identical output with this PR as expected.

Seems like sdpa performs something different, need to investigate more on this.

I'll let you know when I found the root cause : ) thanks!