Add GSM8K evaluation script and AWQ+FP8 results

[rtj1]

This PR adds GSM8K evaluation results for AWQ+FP8 quantization as requested in #2305.

What's included

RESULTS.md - Evaluation results comparing FP8_DYNAMIC vs FP8_BLOCK quantization schemes on Meta-Llama-3-8B-Instruct

Results

Scheme	Strict Match	Flexible Extract
FP8_DYNAMIC	76.42%	76.19%
FP8_BLOCK	75.21%	74.98%

• Model: Meta-Llama-3-8B-Instruct
• Hardware: 8x NVIDIA A100-SXM4-80GB
• FP8_DYNAMIC outperforms FP8_BLOCK by ~1.2% on strict match

Evaluation command

lm_eval \
  --model hf \
  --model_args pretrained=<model_path>,dtype=auto \
  --tasks gsm8k \
  --batch_size 16 \
  --output_path <output_dir>

Note: batch_size=16 is important — the default auto picks 1, significantly increasing evaluation time.

Model Checkpoints (from @HDCharles)

• FP8_DYNAMIC: https://huggingface.co/nm-testing/Meta-Llama-3-8B-Instruct-awq-asym-fp8-dynamic
• FP8_BLOCK: https://huggingface.co/nm-testing/Meta-Llama-3-8B-Instruct-awq-asym-fp8-block

[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 @rtj1, 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 new evaluation script and comprehensive results for AWQ+FP8 quantization on the GSM8K benchmark, specifically targeting the Qwen2.5-0.5B-Instruct model. It demonstrates the superior performance of the FP8_DYNAMIC scheme over FP8_BLOCK in terms of accuracy preservation, while maintaining efficient model compression. The changes aim to provide clear, reproducible benchmarks for quantized models.

Highlights

• New GSM8K Evaluation Script: A new Python script, gsm8k_eval.py, has been added to facilitate benchmarking of quantized models on the GSM8K dataset using the lm-eval harness. This script includes a critical batch_size=16 setting to ensure efficient evaluation, preventing the default auto setting from causing excessively long runtimes.
• Comprehensive AWQ+FP8 Results Document: A new RESULTS.md file is introduced, providing detailed quantization and evaluation results. It covers the Qwen2.5-0.5B-Instruct model, tested with FP8_DYNAMIC and FP8_BLOCK schemes on a Google Colab L4 GPU, including model sizes, evaluation metrics, and a full reproducibility workflow.
• Key Performance Findings for FP8 Quantization: The evaluation reveals that FP8_DYNAMIC quantization significantly outperforms FP8_BLOCK in strict match accuracy on GSM8K (22.67% vs 17.97%). Both schemes achieve a similar compression ratio of approximately 1.2x, reducing the model size from 1.1GB to 0.92GB.
• Enhanced Reproducibility: To ensure reproducibility, the quantized models (FP8_DYNAMIC and FP8_BLOCK) have been uploaded to HuggingFace. The RESULTS.md file also provides a step-by-step bash workflow, including commands to set up and run the quantization and evaluation processes.

🧠 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.

Changelog

• examples/awq/RESULTS.md
  • Added a new markdown file documenting the AWQ + FP8 quantization results for Qwen2.5-0.5B-Instruct.
  • Includes a summary, quantization metrics (size, files), GSM8K evaluation results (Strict Match, Flexible Extract), HuggingFace model links, setup instructions, evaluation details, reproducibility workflow, and a recommendation to use FP8_DYNAMIC.
• examples/awq/gsm8k_eval.py
  • Added a new Python script for evaluating quantized models on the GSM8K dataset.
  • Utilizes lm-eval and sets a fixed batch_size=16 to optimize evaluation runtime, avoiding the slow default auto setting.
  • Takes the model path as a command-line argument and saves results to a dynamically named output directory.

Activity

• The pull request addresses and closes issue Evaluate AWQ + FP8 Example #2305, which requested GSM8K evaluation results for AWQ+FP8 quantization.
• The author rtj1 has added two new files: examples/awq/RESULTS.md and examples/awq/gsm8k_eval.py.
• The author has requested a review or attention from HDCharles via a cc mention in the PR description.

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 introduces a valuable evaluation script for GSM8K and provides detailed results for AWQ+FP8 quantization in RESULTS.md. The documentation is clear and the script is a useful addition for benchmarking. I have a couple of suggestions to enhance the robustness of the reproduction steps and the new evaluation script, mainly by making a shell command more specific and adding input validation to the Python script for better error handling.

[gemini-code-assist]

Using a broad wildcard like *.py with sed -i can be risky, as it might unintentionally modify other Python files in the directory. It would be safer and more robust to use a more specific pattern that targets only the intended example scripts.

Suggested change

	sed -i 's/meta-llama\/Meta-Llama-3-8B-Instruct/Qwen\/Qwen2.5-0.5B-Instruct/g' *.py
	sed -i 's/meta-llama\/Meta-Llama-3-8B-Instruct/Qwen\/Qwen2.5-0.5B-Instruct/g' fp8_*_llama_example.py

[HDCharles]

So we're looking to evaluate the actual models the examples are generating. I'll run evals using your PR and we can go from there

[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

[rtj1]

Thanks for taking a look, @HDCharles!

The quantized models are uploaded to HuggingFace:

• FP8_DYNAMIC: https://huggingface.co/rtj1/Qwen2.5-0.5B-AWQ-FP8-Dynamic
• FP8_BLOCK: https://huggingface.co/rtj1/Qwen2.5-0.5B-AWQ-FP8-Block

I've also fixed the quality checks - all tests passing now. Let me know if you see any issues with the models or need different configs for the evaluation.

[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

[HDCharles]

looks like you edited a ton of files. I would undo the last commit and reinstall llm-compressor to get the correct version of ruff pip install -e.[dev] and try make style and make quality again.

[HDCharles]

ok here is my evaluation, see https://github.com/vllm-project/llm-compressor/pull/2347/changes, can you finalize this PR using those numbers

checkpoints

https://huggingface.co/nm-testing/Meta-Llama-3-8B-Instruct-awq-asym-fp8-block/tree/main/
https://huggingface.co/nm-testing/Meta-Llama-3-8B-Instruct-awq-asym-fp8-dynamic/tree/main/

[rtj1]

Updated with Llama-3-8B evaluation results from PR #2347.

Reinstalled llm-compressor with dev dependencies (pip install -e.[dev]) and reran code formatting as requested. All quality checks now passing.

The PR now includes your official Llama-3-8B-Instruct evaluation results showing FP8_DYNAMIC achieves 76.42% vs FP8_BLOCK's 75.21% on GSM8K strict matching.

[brian-dellabetta]

no need to call this out in the changes, just in the PR summary

[brian-dellabetta]

I think we should remove this file in favor of directly calling lm_eval as you show in the markdown file.

[rtj1]

Thanks for the feedback @brian-dellabetta! I've addressed both points:

• Removed Closes #2305 from RESULTS.md
• Removed gsm8k_eval.py and updated docs to use lm_eval directly

Latest commit: 002b77a

[HDCharles]

Suggested change

FP8_DYNAMIC wins by ~1.2% on strict matching. Both achieve similar performance on flexible extraction.

this seems outdated?

[rtj1]

Done, removed. Thanks!

[HDCharles]

update the PR description for the targeted model