Skip to content

Conversation

@rohansjoshi
Copy link
Contributor

Summary:
Added option to use MSE range setting algorithm. This algorithm does a linear grid search over scales and selects those which minimizes mean squared error (see the line_search method in the class PerChannelParamsObserver). This method is applied for quantizing weights per channel. Accuracy is still poor, but somewhat better than using MinMax.

On wikitext task, with grid size 200:

Model Name max_seq_len ptq word_perplexity
Llama 3.2-1B Instruct 128 16a4w 2367107
Llama 3.2-1B Instruct 128 16a4w_block 5523977
Llama 3.2-1B Instruct 128 8a8w 501663

Reviewed By: cccclai

Differential Revision: D77055545

@rohansjoshi rohansjoshi requested a review from cccclai as a code owner June 23, 2025 18:05
@pytorch-bot
Copy link

pytorch-bot bot commented Jun 23, 2025

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/11857

Note: Links to docs will display an error until the docs builds have been completed.

❌ 2 Cancelled Jobs

As of commit 83dc888 with merge base 2f55193 (image):

CANCELLED JOBS - The following jobs were cancelled. Please retry:

This comment was automatically generated by Dr. CI and updates every 15 minutes.

@facebook-github-bot facebook-github-bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Jun 23, 2025
@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D77055545

@github-actions
Copy link

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.

rohansjoshi added a commit to rohansjoshi/executorch that referenced this pull request Jul 4, 2025
Summary:

Added option to use MSE range setting algorithm. This algorithm does a linear grid search over scales and selects those which minimizes mean squared error (see the line_search method in the class PerChannelParamsObserver). This method is applied for quantizing weights per channel.

On wikitext task:
| Model Name | max_seq_len | ptq | range_setting | word_perplexity
|----------|----------|----------|-----------|-----------|
| Llama 3.2-1B Instruct  | 1024   | 16a4w |  MinMax | 184.06065814138435 |
| Llama 3.2-1B Instruct  | 1024   | 16a4w |  MSE (grid size 200) | 292.199408962586 |

Reviewed By: cccclai

Differential Revision: D77055545
@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D77055545

rohansjoshi added a commit to rohansjoshi/executorch that referenced this pull request Jul 6, 2025
Summary:

Added option to use MSE range setting algorithm. This algorithm does a linear grid search over scales and selects those which minimizes mean squared error (see the line_search method in the class PerChannelParamsObserver). This method is applied for quantizing weights per channel.

On wikitext task:
| Model Name | max_seq_len | ptq | range_setting | word_perplexity
|----------|----------|----------|-----------|-----------|
| Llama 3.2-1B Instruct  | 1024   | 16a4w |  MinMax | 184.06065814138435 |
| Llama 3.2-1B Instruct  | 1024   | 16a4w |  MSE (grid size 200) | 292.199408962586 |

Reviewed By: cccclai

Differential Revision: D77055545
Summary:
Pull Request resolved: pytorch#11857

Added option to use MSE range setting algorithm. This algorithm does a linear grid search over scales and selects those which minimizes mean squared error (see the line_search method in the class PerChannelParamsObserver). This method is applied for quantizing weights per channel.

On wikitext task:
| Model Name | max_seq_len | ptq | range_setting | word_perplexity
|----------|----------|----------|-----------|-----------|
| Llama 3.2-1B Instruct  | 1024   | 16a4w |  MinMax | 184.06065814138435 |
| Llama 3.2-1B Instruct  | 1024   | 16a4w |  MSE (grid size 200) | 292.199408962586 |

Reviewed By: cccclai

Differential Revision: D77055545
@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D77055545

@facebook-github-bot facebook-github-bot merged commit 6669637 into pytorch:main Jul 7, 2025
101 of 105 checks passed
Tanish2101 pushed a commit to Tanish2101/executorch that referenced this pull request Jul 9, 2025
Differential Revision: D77055545

Pull Request resolved: pytorch#11857
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. fb-exported

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants