[V1][Mamba] - Enable V1 by default for Mamba Models

[Josephasafg]

Purpose

This PR enables V1 by default to Mamba models so they won't fall back to V0. This PR needs to have this PR and this PR that enables full cuda graph support by default merged first, so users won't have to specify VLLM_ATTENTION_BACKEND=FLASHINFER when they start vLLM.

Test Plan

Test Result

---

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.

[Josephasafg]

@tdoublep @heheda12345 Once we get PR and this PR merged, we should probably enable Mamba models by default to V1.

[gemini-code-assist]

Code Review

This pull request aims to enable the V1 engine by default for Mamba models by removing the check that was previously falling back to the V0 engine. While the change is correct in principle, it introduces a critical issue where prefix caching is enabled by default for these models, which they do not support, leading to a crash. A fix is required to adjust the default V1 arguments for Mamba-like models.

[tdoublep]

This change will enable V1 by default for all models that use mamba1, mamba2, minimax linear attention and short conv layers.

For mamba2, we definitely don't want to do this until we first enable cudagraph_mode=FULL_AND_PIECEWISE as the default, because otherwise the performance drop between V0 and V1 is very big.

[tdoublep]

Let's also merge this tiny one first (otherwise user with get a crash using default vllm serve)

[heheda12345]

@tdoublep what about merging the two PRs in parallel? There's no merge conflict between them

[tdoublep]

LGTM

[tdoublep]

@heheda12345 We can. If this one goes first, there will be a short time inbetween when vllm crashes by default for these models. But it's OK - both PRs can be merged in next hours.

[Josephasafg]

@tdoublep I need to set V0 in test_hybrid.py and remove V1 env set. Fixing

[Josephasafg]

I'll wait for this PR to be merged first as my tests kind of depend on them

[heheda12345]

@tdoublep fine for this PR but I think this line makes vLLM not that plugable to new models.

[heheda12345]

@tdoublep Can you create a new PR to update the doc for recent status and a simple guideline for contributing new mamba models?

[Josephasafg]

@heheda12345 @tdoublep I fixed some tests. Some tests are, I believe, only suited for V0 like

• test_chunked_prefill
• test_models_preemption_recompute

[tdoublep]

@Josephasafg Thanks - the tests are passing now. When we remove V0 code we can either decide to adapt those 2 tests to V1 or drop them if they don't make sense.

[tdoublep]

@tdoublep Can you create a new PR to update the doc for recent status and a simple guideline for contributing new mamba models?

@heheda12345 Sure, I will do that. Are there any examples of guidelines for contributing other models that I could use as a reference?

[Josephasafg]

@Josephasafg Thanks - the tests are passing now. When we remove V0 code we can either decide to adapt those 2 tests to V1 or drop them if they don't make sense.

Sounds good.