Feat/Add Compactor press

[vnchari]

PR description

Adds Compactor Press, from https://arxiv.org/abs/2507.08143. It does quite well, and ends up being reasonably lightweight (about the same inference runtime as SnapKV). So far the code only supports prefill (otherwise we would need to unrotate keys, and keep a long buffer of queries, which slows down decoding considerably), so I modified the test-runner to skip CompactorPress when testing the decoder press.

Please let me know if there are any issues! Thanks!

Compression Ratio	0.25	0.5	0.75	0.95
Llama 3.1-8B	94.0	87.7	77.8	63.5
Qwen3-8B	92.6	85.0	72.3	51.7

Checklist

Before submitting a PR, please make sure:

• [✅] Tests are working (make test)

• [✅] Code is formatted correctly (make style, on errors try fix with make format)

• [✅] Copyright header is included

• [✅] All commits are signed-off using git commit -s (it says I didn't, but the commits show that I did. not sure what's wrong here)

• [✅] (new press) mypress_press.py is in the presses directory

• [✅] (new press) MyPress is in __init__.py

• [✅] (new press) README.md is updated with a 1 liner about the new press in the Available presses section

• [✅] (new press) New press is in the default_presses list in tests/default_presses.py

• [✅] (new press) A docstring is provided that follows the same structure as the existing ones

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[maxjeblick]

Thanks a lot for this PR, the method looks quite interesting!
We will go through your paper and code this week and report back.

As for the DCO error, all Nvidia OS repositories require commits to be signed off.

[maxjeblick]

Thanks a lot for the PR, it was a real pleasure reading the paper and code!

Regarding the code, it looks good to me overall; I left a few comments.

There's one suggestion to allow for additional experiments:
Since you are blending the z-normalized scores of two methods, it would make sense to introduce two independent scorer presses (one for l_scores, one for attn_scores) and define your press as a wrapper press that blends both base presses.
By this, it would be possible to combine the outlier detection with other presses defined in the library. We (the maintainers) can also work on this potential feautre in a subseuqent PR if you prefer to keep the press as is.

[vnchari]

I'm glad you enjoyed the paper :) and thanks for the feedback. The paper should indeed say that we center the key states, and I will update it to do so. As you suggested, I split the press into two presses and added some clarifying comments + tests. Also, I have some code ready to go for "calibrated" compression as described in the paper (for any ScorerPress). Do you think it would be best open a new PR for that, or add it here?

Please let me know if there are any issues! Thanks!

[maxjeblick]

Also, I have some code ready to go for "calibrated" compression as described in the paper (for any ScorerPress). Do you think it would be best open a new PR for that, or add it here?

This sounds great! IMO, it is best to open another PR implementing "calibrated" compression.

[maxjeblick]

/ok to test a391627

[maxjeblick]

LGTM, thanks a lot!

Regarding DCO:

• You should be able to sign off the missing commits.
• Otherwise, feel free to open a new PR with the same code (stashed to a single, signed commit)

[vnchari]

Done!

[maxjeblick]

/ok to test 6552dff