Support vLLM 0.12.0

[jiqing-feng]

The latest vllm changed the guided_decoding to structured_outputs in this PR 22772. We also need to update this parameter in the vllm client and vllm serve.

Hi @qgallouedec . Would you please review this PR? Thanks!

[kashif]

we will need to make sure that we require the version of vllm that supports this feature

[jiqing-feng]

we will need to make sure that we require the version of vllm that supports this feature

Sure, we can wait for the vllm release and then update the vllm version in setup in this PR.

[kashif]

you will need to update this: https://github.com/huggingface/trl/blob/main/setup.cfg#L67

[jiqing-feng]

you will need to update this: https://github.com/huggingface/trl/blob/main/setup.cfg#L67

Yes, exactly. I will update it once vllm 10.3.0 released.

[qgallouedec]

for ref: #4122 will pin vLLM

[ed-pai]

Issue tracking this PR: #4617

[jiqing-feng]

Hi @kashif @qgallouedec . I have updated this PR. Please review it. Thanks!

[ed-pai]

Thanks @jiqing-feng! BTW, it looks like you're pinning to <v0.12.0. What's the blocker now for v0.12.0? It was released last week.

[jiqing-feng]

Thanks @jiqing-feng! BTW, it looks like you're pinning to <v0.12.0. What's the blocker now for v0.12.0? It was released last week.

I just kept the original maximum version v0.12.0, it's not my change.

[ed-pai]

My understanding is that TRL already worked with v0.11.2. This fix is required for support of v0.12.0.
It doesn't make sense to not allow that version in the pyproject.toml ("vllm>=0.10.3,<0.12.0").

@qgallouedec, wanted to ensure this PR was on your radar. Thanks!

[qgallouedec]

Thanks for the PR. Indeed, TRL supports 0.10.2 - 0.11.2, but not 0.12, and this PR should allow to support 0.12. Ideally we would like to continue supporting 0.10.2 - 0.11.2, I'm not sure if this PR allows this, I'll need to check. I'll review it in details very soon.

[qgallouedec]

0.10.3 is not a valid vllm version.

[qgallouedec]

StructuredOutputsParams is not in vLLM 0.10.2. We should have a version check to import GuidedDecodingParams when vLLM version is 0.10.2, and StructuredOutputsParams otherwise. Also, to avoid breaking change, let's have 1 version of deprecation for vllm_guided_decoding_regex

[qgallouedec]

LGTM, thanks! I modified a bit to ensure backward compatibility

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.