[Transform] [Quantization] Add transforms to compressed tensors

[kylesayrs]

Online Hadamard Rotations

Purpose

• Support online linear transforms for models compressed with CompressedTensors

Changes

Added Transforms Weight Loading

• Implemented SharedWeightParameter
  • This class implements a registry where tensors can be shared. Loading with shared tensors is required for models with many transforms such as QuIP
• Before weight loading, a HadamardTransform module is attached to linear layers to load transform weights which utilizes SharedWeightParameter in order to load weight partitions as separate tensors

Added Transforms Apply

• Add CompressedTensorsLinearTransformMethod which wraps CompressedTensorsLinearMethod and UnquantizedLinearMethod adds input and output transforms to either side of the original apply method
  • Unfortunately, because [Core] Support weight_loader_v2 for UnquantizedLinearMethod #23036 has not landed, we must use a hack to switch back to weight_loader_v1 if the method being wrapped is the UnquantizedLinearMethod
  • This class has template functionality for selecting which kernel to use when applying transforms

Misc Changes

• Add transform config parsing to CompressedTensors
• Define _shard_id_as_int on BasevLLMParameter. This is so that its implementation can be shared with SharedWeightParameter
• Add test utility calculate_prompt_perplexity for checking model coherence
• Added weight_loader_v1 to QKVCrossParallelLinear to support hack described above (and [Core] Support weight_loader_v2 for UnquantizedLinearMethod #23036)

Testing

• Added test_compressed_tensors_transforms_perplexity tests for SpinQuantR1R2R4 and QuIP
• Locally tested unquantized QuIP and SpinQuantR1R2R4 models

[gemini-code-assist]

Code Review

This pull request introduces support for transforms within the compressed tensors quantization framework. The changes primarily involve updating CompressedTensorsConfig to manage transform configurations and modifying the linear method to apply these transforms during the forward pass. My review has identified two critical issues that need to be addressed. First, the logic for creating transform factories incorrectly overwrites the factories dictionary within a loop, which will lead to incorrect behavior when multiple transform schemes are present. Second, there is a signature mismatch in the call to the newly added is_match function, which will cause a TypeError at runtime.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @kylesayrs.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[yewentao256]

Let's run CI and see if it is correct

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @kylesayrs.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork