[V1] [Hybrid] Disable prefix caching by default for hybrid or mamba-based models

[tdoublep]

Purpose

We would like to enable V1 by default for hybrid models (or models based on "mamba" layers, where "mamba" is a stand-in for: mamba2, mamba2, linear_attention or short_conv). However, these models do not yet support prefix caching. This PR will disable prefix caching by default for these models, ensuring that user does not experience crash using default vllm serve ....

This is just a user experience improvement until we enable prefix caching - we are aiming to put up a first PR for this later this week.

cc @heheda12345 @asafgardin

Test Plan

n/a

Test Result

n/a

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[gemini-code-assist]

Code Review

This pull request disables prefix caching for Mamba-based and hybrid models to prevent crashes, as this feature is not yet supported for them. This is a valuable user experience improvement. My review includes a suggestion to refine the implementation. Instead of unconditionally disabling the feature, I recommend checking if the user has explicitly enabled it and then issuing a warning before disabling. This approach enhances clarity for the user and aligns better with existing configuration handling practices in the codebase.

[Josephasafg]

LGTM

[Josephasafg]

@tdoublep Should we also update v1_guide? since the users dont need to disable prefix caching after this change

[tdoublep]

@Josephasafg Good catch thanks - I have updated the language accordingly.