[Misc] Split up pooling tasks

[DarkLight1337]

As part of #10674, split up the embedding task into the following tasks for pooling models:

• Embedding (embed), with default pooler pooling_type=LAST, normalize=True
  • Usually one output per prompt
• Classification (classify), with default pooler pooling_type=LAST, softmax=True
  • Usually one output per prompt
  • The scoring module will be automatically loaded from sentence-transformers.
  • Other models should register a separate architecture so that the model contains the correct modules.
• Scoring (scoring)
  • Technically an alias for Classification, used for cross-encoder models
• Reward Modeling (reward), with default pooler pooling_type=ALL
  • Usually one output per token per prompt
  • If the reward model has any additional modules compared to the text generation model, a separate architecture should be registered so that the model contains the correct modules.
  • To avoid peiyi9979/math-shepherd-mistral-7b-prm conflicting with other LlamaForCausalLM pooling models, we will remove the pooler default from the text generation model. Since users already have to override step_tag_id etc. anyway to use that model, this should not be a major breaking change.

Backwards compatibility

• embedding remains available as an alias to embed.
• If --task embedding is passed but the requested pooling model is not an embedding model, a deprecation warning is emitted.
• Later, we will make --task embedding exclusive to embedding models.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
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 do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[DarkLight1337]

cc @russellb see if these new doc additions look good to you!

[DarkLight1337]

@Isotr0py would be great if you could review this as well!

[Isotr0py]

Overall LGTM. Just a comment about the score task.

[Isotr0py]

LGTM now!