sparseml

sparseml copied to clipboard

This code updates an existing but unused model analyzer (AnalyzerModule) object that computes forward FLOPs, parameters, prunable parameters, and zeroed parameters model-wide. Note that this is somewhat redundant with the...

ohaijen

Code that I used to evaluate Llama-2-7b model on CNN/DailyMail dataset with the Rouge score.

natuan

## Short-term Work Issues or PRs that the NM team are planning to tackle for this quarter: * SparseGPT one-shot pruning for Transformers * [x] Support for text generation LLMs...

mgoin

The PR adds support for utilizing HistogramObserver from PyTorch which computes the min/max values for quantization by minimizing quantization error. The implementation has been tested on CodeLlama and Llama-2 models.

abhinavnmagic

rahul-tuli

Add support for loading Transformers models without specifying task attributes. This is especially useful for exporting models for embedding extraction. This current accessed by `"model"` or `"base"` task - I'm...

mgoin

Bumps [scikit-learn](https://github.com/scikit-learn/scikit-learn) from 0.24.2 to 1.0.1. Release notes Sourced from scikit-learn's releases. scikit-learn 1.0.1 We're happy to announce the 1.0.1 release with several bugfixes: You can see the changelog here:...

dependabot[bot]

RecursionError when converting LlaMa model to ONNX

5

**Describe the bug** `RecursionError: maximum recursion depth exceeded while getting the str of an object` **Expected behavior** I want to the convert a LlaMa model into ONNX and then benchmark...

luuyin

Additional tests to ensure Top-KAST is working as intended. Bugfix: when computing weight decay for the backwards-only weights (set B in the paper), the multiplier should be proportional to 1/(the...

ohaijen

Not ready for prime time, but it does work in making LLM export much more memory efficient and faster

mgoin