chore: update transformers to support Qwen3 architecture

[aryasaatvik]

Related Issue

Adds support for Qwen3 model architecture.

Summary

This PR updates the transformers dependency to ensure compatibility with Qwen3 models.

Changes:

• Updated transformers version constraint in pyproject.toml to >=4.51.3,<=5.0
• Regenerated poetry.lock with updated dependencies

Why this is needed:

• Enables users to serve Qwen3 models
• Keeps dependencies up to date with latest model architectures

Checklist

• I have read the CONTRIBUTING guidelines.
• I have added tests to cover my changes.
• I have updated the documentation (docs folder) accordingly.

Additional Notes

This is a dependency update that maintains backward compatibility.

[greptile-apps]

PR Summary

Updates transformers dependency in libs/infinity_emb/pyproject.toml to support Qwen3 architecture by raising minimum version to 4.51.3.

• Updated transformers dependency from >=4.47.0,<=5.0 to >=4.51.3,<=5.0 while maintaining backward compatibility
• Missing test coverage for Qwen3 model support despite being a checklist item
• No documentation updates in docs folder to reflect new model support
• Poetry.lock updated but changes not included in the review context

_{1 file reviewed, no comments
Edit PR Review Bot Settings | Greptile}

[Abhi011999]

can we please merge this @michaelfeil
qwen3 is now the current SOTA

[zengqingfu1442]

Only updating transformers can not solve the problem and cannot still deploy Qwen3-reranker-8B. see #611 (comment)

[aryasaatvik]

Only updating transformers can not solve the problem and cannot still deploy Qwen3-reranker-8B. see #611 (comment)

try using tomaarsen/Qwen3-Reranker-4B-seq-cls

[zengqingfu1442]

Only updating transformers can not solve the problem and cannot still deploy Qwen3-reranker-8B. see #611 (comment)

try using tomaarsen/Qwen3-Reranker-4B-seq-cls

What is the difference between tomaarsen/Qwen3-Reranker-8B-seq-cls and official Qwen/Qwen3-reranker-8B? thanks.

[Abhi011999]

The difference is mentioned here: https://huggingface.co/Qwen/Qwen3-Reranker-0.6B/discussions/3
They also implemented some hacky method to integrate with vllm: vllm-project/vllm#19260

Apparently, the original Qwen/Qwen3-Reranker-8B model's both embedder and reranker support Qwen3ForCausalLM architecture.

I don't think infinity has a way to support the reranker as of now through Qwen3ForSequenceClassification, they already implemented it in vllm. I am thinking if there's a way to support it here too.

[aryasaatvik]

The difference is mentioned here: huggingface.co/Qwen/Qwen3-Reranker-0.6B/discussions/3 They also implemented some hacky method to integrate with vllm: vllm-project/vllm#19260

Apparently, the original Qwen/Qwen3-Reranker-8B model's both embedder and reranker support Qwen3ForCausalLM architecture.

I don't think infinity has a way to support the reranker as of now through Qwen3ForSequenceClassification, they already implemented it in vllm. I am thinking if there's a way to support it here too.

I was able to get it working on modal https://gist.github.com/aryasaatvik/6f868d5f064b25e41cf1e20018721dca

[zengqingfu1442]

The difference is mentioned here: https://huggingface.co/Qwen/Qwen3-Reranker-0.6B/discussions/3 They also implemented some hacky method to integrate with vllm: vllm-project/vllm#19260

Apparently, the original Qwen/Qwen3-Reranker-8B model's both embedder and reranker support Qwen3ForCausalLM architecture.

I don't think infinity has a way to support the reranker as of now through Qwen3ForSequenceClassification, they already implemented it in vllm. I am thinking if there's a way to support it here too.

The vllm 0.9.2 already support to deploy Qwen3-Reranker-8B and Qwen3-Embedding-8B.

[michaelfeil]

A update of the transformer libary will break all

• onnx-optimum
• bettertransformer dependencies

There is a large undertaking to migrate these libaries. Migrating bert from bettertransformer to e.g. torch + compile will lead to a 50% performance loss in throughput and latency.