Update to Transformers v4.56.2

[hmellor]

Supersedes #24043

[gemini-code-assist]

Code Review

This pull request updates the transformers dependency to version 4.56.1 in several requirement files. The changes are straightforward, involving a simple version number update across multiple files.

[DarkLight1337]

Looks like HF backend for Idefics3 is broken by this upgrade: https://buildkite.com/vllm/ci/builds/30368/steps/canvas?sid=01993997-f436-4816-b6bb-ac9f68f04cc2

[hmellor]

This test is not currently run on main:

tests/models/language/generation/test_hybrid.py::test_models[5-64-yujiepan/mamba2-codestral-v0.1-tiny-random] SKIPPED (`transformers==4.55.2` installed, but `transformers>=4.55.3` is required to run this model.)

Even when upgrading to the most similar version, v4.55.3, this test fails. @tdoublep I see you added this in #23936, did this test pass for you in your development environment?

[hmellor]

This test is not currently run on main:

tests/models/multimodal/processing/test_common.py::test_processing_correctness[1.0-32-0.3-zai-org/GLM-4.5V] SKIPPED (`transformers==4.55.2` installed, but `transformers>=4.56` is required to run this model.)

When upgrading to v4.56.0 or v4.56.1, this test still fails with:

>               if metadata["total_num_frames"] != len(video_array):
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
E               TypeError: 'NoneType' object is not subscriptable

vllm/model_executor/models/glm4_1v.py:1132: TypeError

@DarkLight1337 I see you labelled this as requiring v4.56 in #23181, has anyone run this test locally?

[hmellor]

The Emu3 test requires a fix on the Transformers side, we'll make a patch for this ASAP

[DarkLight1337]

No, I marked it as v4.56 only because the model isn't in v4.55.0 release

[hmellor]

Ok, the zai-org/GLM-4.5V failure appears to be a problem on the vLLM side. @zRzRzRzRzRzRzR could you look into this?

[tdoublep]

@hmellor I did test it in my local dev environment some time ago (when I put up that PR) and it passed. But I confirm that it doesn't pass on main right now. I will try to see if I can figure out what changed. We are also skip the test for now if it's blocking something.

[DarkLight1337]

Ok, the zai-org/GLM-4.5V failure appears to be a problem on the vLLM side.

@Isotr0py is this patched by #24161 (comment)?

[Isotr0py]

is this patched by #24161 (comment)?

Yes, and the corresponding fix should be quite similar to Qwen3-VL's video processing implementation. So I will fix it as long as we addressed the Qwen3-VL processing test failure in #24727 (comment)

[hmellor]

Thanks @tdoublep, it would be nice if we could not postpone the fix, but if if's too difficult to solve quickly I can change the reason the test skips so that this PR can be merged.

[tdoublep]

@hmellor I must have made a mistake previously - I don't see how it could have been working in my local dev environment. The model is broken on transformers because the n_groups=8 parameter is ignored in the RMSNorm. I've put up a PR to fix that here: huggingface/transformers#40861

I can confirm that with that fix the test is passing:

python -mpytest tests/models/language/generation/test_hybrid.py::test_models[5-64-yujiepan/mamba2-codestral-v0.1-tiny-random] 

[hmellor]

Thanks for the PR @tdoublep! I'll see what I can do about getting it expedited so that we can upgrade to v4.56.2 ASAP

[Isotr0py]

Ok, the zai-org/GLM-4.5V failure appears to be a problem on the vLLM side.

The processor test should be fixed by #24822.

[DarkLight1337]

Can you update this PR? The tests should be pass after #25644

[DarkLight1337]

Hmm, maybe we still need the changes in transformers backend?

[hmellor]

Seems like it, I'll re-add the changes in a more polished way next week

[hmellor]

I've seen these errors in other PRs that have recently been updated from main