Improve the output precision of embedding models

[noooop]

Convert the embedding output to float32, otherwise precision will be lost significantly

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[noooop]

@DarkLight1337

I tested 46 models (including skipped model) participating in ci test_embed_models_mteb.

Among them, 45 could pass the test using auto dtype strategy.

BAAI/bge-code-v1 must use float32 to pass the test. Using hybrid dtype also cannot pass the test.

So we don't need to use hybrid dtype.

Can we merge this pr first.

Identify which step reduced precision when we refactored the pooler and PoolingOutput.

[DarkLight1337]

Sounds good, thanks!

[noooop]

@DarkLight1337

tpu-v1-test hanging

#19115

We definitely won't wait for it to end automatically

please manually close and release the CI resources