Added GLM Modeling

[phaelon74]

SUMMARY:
Without a model file for GLM, only experts activated from a calibaration sample, are used. This means accuracy will drop quite a bit if the dataset is not robust and even with a robust dataset, I highly doubt all experts are being hit and pulled from as needed.

With this model file, it should cycle through all experts like Qwen does with it's model files, etc.

TEST PLAN:
I used LLM_Compressor to quant GLM4.7 and it succeeded without issue. Equally I quanted a llama3 model as well, to make sure it didn't break any other flows.

[github-actions]

👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review.

Note: This is required to complete the testing suite, please only add the label once the PR is code complete and local testing has been performed.

[gemini-code-assist]

Summary of Changes

Hello @phaelon74, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request introduces a specialized calibration module for GLM4 Mixture-of-Experts (MoE) models within the llmcompressor framework. The primary goal is to enhance the accuracy and robustness of quantized GLM4 models by ensuring that all experts within the MoE architecture are thoroughly calibrated, rather than just those activated by a limited calibration sample. This is achieved by temporarily routing all tokens to all experts during the calibration phase, while still respecting routing weights for the final output.

Highlights

• GLM4 MoE Calibration Module: A new CalibrationGlm4MoeMoE module has been introduced to enable proper calibration for GLM4 Mixture-of-Experts (MoE) models. This module ensures that during calibration, all tokens are sent to all experts, allowing for comprehensive quantization statistics collection, even for experts not activated by standard routing.
• Enhanced Calibration Accuracy: This change addresses a limitation where, without a dedicated model file, only experts activated by a calibration sample were used, potentially leading to reduced accuracy. The new module ensures all experts are properly calibrated, improving the overall robustness and accuracy of quantized GLM4 models.
• Integration and Testing: The new GLM4 MoE calibration module is integrated into the llmcompressor.modeling package. Testing confirmed successful quantization of GLM4.7 and verified that existing quantization flows for other models like Llama3 remain unaffected.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request adds support for GLM MoE model calibration by introducing CalibrationGlm4MoeMoE. The implementation correctly handles passing all tokens to all experts during calibration, which is crucial for accurate quantization. My feedback includes a suggestion to refactor the forward method to improve maintainability and a strong recommendation to add unit tests for this new functionality to ensure its correctness and prevent future regressions.

[mratsim]

Nice, you frontrunned me!

I was tired of waiting for Transformers v5 before doing proper all experts calibration of GLM-4.7 / GLM-4.5-Air / GLM-4.6V (see #2036).

Not doing so can have a dramatic impact and many models out there just use raw LLM compressors (mine included): https://avtc.github.io/aquarium-side-by-side/

[mratsim]

By the way, does this PR work on the vision models as-is like GLM-4.6V?

[phaelon74]

By the way, does this PR work on the vision models as-is like GLM-4.6V?

It should handle it yes, schematically, but I can run it through the test script later tonight and tell you for certain.

[phaelon74]

I believe this is now ready for formal review, after having resolved the issues and added a testing script with results. This is my first PR, so if I missed anything, apologies.

[dsikka]

Thank you! Would you mind adding an example which uses the new modeling code?

[phaelon74]

Thank you! Would you mind adding an example which uses the new modeling code?

As in my quant script, or a model quanted with my quant script using the new modeling file?

[dsikka]

Thank you! Would you mind adding an example which uses the new modeling code?

As in my quant script, or a model quanted with my quant script using the new modeling file?

A quant script is fine. Thanks!

[phaelon74]

Thank you! Would you mind adding an example which uses the new modeling code?

As in my quant script, or a model quanted with my quant script using the new modeling file?

A quant script is fine. Thanks!

Here's my quant script: https://github.com/phaelon74/PhaeDawg-QuantScripts_Compressed-Tensors/blob/main/GLM-4.7/glm4.7_W4A16_GS32.py

[dsikka]

Thank you! Would you mind adding an example which uses the new modeling code?

As in my quant script, or a model quanted with my quant script using the new modeling file?

A quant script is fine. Thanks!

Here's my quant script: https://github.com/phaelon74/PhaeDawg-QuantScripts_Compressed-Tensors/blob/main/GLM-4.7/glm4.7_W4A16_GS32.py

Do you mind adding it as example to this PR?

[phaelon74]

Thank you! Would you mind adding an example which uses the new modeling code?

As in my quant script, or a model quanted with my quant script using the new modeling file?

A quant script is fine. Thanks!

Here's my quant script: https://github.com/phaelon74/PhaeDawg-QuantScripts_Compressed-Tensors/blob/main/GLM-4.7/glm4.7_W4A16_GS32.py

Do you mind adding it as example to this PR?

Jused Added it. Hope I did it right.

[dsikka]

Thank you! Would you mind adding an example which uses the new modeling code?

As in my quant script, or a model quanted with my quant script using the new modeling file?

A quant script is fine. Thanks!

Here's my quant script: https://github.com/phaelon74/PhaeDawg-QuantScripts_Compressed-Tensors/blob/main/GLM-4.7/glm4.7_W4A16_GS32.py

Do you mind adding it as example to this PR?

Jused Added it. Hope I did it right.

Thank you! Do you mind addressing the quality issues?

[brian-dellabetta]

Thanks for the contribution! Leaving a few comments for now, we can iterate over this

[phaelon74]

Thank you! Would you mind adding an example which uses the new modeling code?

As in my quant script, or a model quanted with my quant script using the new modeling file?

A quant script is fine. Thanks!

Here's my quant script: https://github.com/phaelon74/PhaeDawg-QuantScripts_Compressed-Tensors/blob/main/GLM-4.7/glm4.7_W4A16_GS32.py

Do you mind adding it as example to this PR?

Jused Added it. Hope I did it right.

Thank you! Do you mind addressing the quality issues?

My apologies, but are you meaning the items @brian-dellabetta called out below, or something else? I'm not sure what quality issues we're specifically talking about. Sorry, very new to this, so clunking my way through it.

[phaelon74]

Hi @phaelon74,

I'd still like to see the GLM example simplified and standardized to match the format of the other example scripts. You can choose to not include either the GPTQ or AWQ examples.

For the two you provided, I don't see any other examples where we differentiate between AWQ/GPTQ. Is there a separate folder for GPTQ? I thought GPTQ had fallen out of favor/not used as much? Thanks again for your engagement and feedback!

There's a lot of mixed messaging/ sometimes incorrect information about GPTQ vs AWQ in the quantization community. GPTQ is our team's preferred algorithm, in large part due to the heavy memory and runtime requirements of AWQ, although both are strong algorithms!

@kylesayrs I grabbed your simplified script, and added the import for GLM4_7 and a simplified MoE layer ignore for the first three dense layers for glm4_7. This should be much closer to the others example scripts.

[phaelon74]

This LGTM once @kylesayrs is happy with the example script. @phaelon74 Thank you for your contribution! Do you have any generated checkpoints that you would be able to share through the HF Hub?

I updated a simplified script so hopefully we're good there.

As for checkpoints, you mean a model quantized in the method aligned and shared on HF? If so, I've got a GS32, GS64, and GS128 of GLM4.7 here made from my script and modeling file: https://huggingface.co/TheHouseOfTheDude/GLM-4.7_Compressed-Tensors/tree/main

[dsikka]

This LGTM once @kylesayrs is happy with the example script. @phaelon74 Thank you for your contribution! Do you have any generated checkpoints that you would be able to share through the HF Hub?

I updated a simplified script so hopefully we're good there.

As for checkpoints, you mean a model quantized in the method aligned and shared on HF? If so, I've got a GS32, GS64, and GS128 of GLM4.7 here made from my script and modeling file: https://huggingface.co/TheHouseOfTheDude/GLM-4.7_Compressed-Tensors/tree/main

For checkpoints, I mean a model produced from this PR. The link you shared doesnt seem to have any safetensors files for the model

[phaelon74]

This LGTM once @kylesayrs is happy with the example script. @phaelon74 Thank you for your contribution! Do you have any generated checkpoints that you would be able to share through the HF Hub?

I updated a simplified script so hopefully we're good there.
As for checkpoints, you mean a model quantized in the method aligned and shared on HF? If so, I've got a GS32, GS64, and GS128 of GLM4.7 here made from my script and modeling file: https://huggingface.co/TheHouseOfTheDude/GLM-4.7_Compressed-Tensors/tree/main

For checkpoints, I mean a model produced from this PR. The link you shared doesnt seem to have any safetensors files for the model

I store them in branches in my main repo on HF:

[dsikka]

This LGTM once @kylesayrs is happy with the example script. @phaelon74 Thank you for your contribution! Do you have any generated checkpoints that you would be able to share through the HF Hub?

I updated a simplified script so hopefully we're good there.
As for checkpoints, you mean a model quantized in the method aligned and shared on HF? If so, I've got a GS32, GS64, and GS128 of GLM4.7 here made from my script and modeling file: https://huggingface.co/TheHouseOfTheDude/GLM-4.7_Compressed-Tensors/tree/main

For checkpoints, I mean a model produced from this PR. The link you shared doesnt seem to have any safetensors files for the model

I store them in branches in my main repo on HF:

Oh perfect, thank you!

[mergify]

The quality checks have failed. Please run make style and make quality under
the root directory to adddress the lint failures. You will need to install the
dev optional install to get the required linting packages:
https://github.com/vllm-project/llm-compressor/blob/main/CONTRIBUTING.md

[dsikka]

Thank you! Please resolve the quality failure

[phaelon74]

Thank you! Please resolve the quality failure

I added back the comment to the import line, should be good to go now.

[mergify]

The quality checks have failed. Please run make style and make quality under
the root directory to adddress the lint failures. You will need to install the
dev optional install to get the required linting packages:
https://github.com/vllm-project/llm-compressor/blob/main/CONTRIBUTING.md

[kylesayrs]

Looks great! Thanks for updating the examples!