[6/n] Replace skip-softmax calibration formula

[kaix-nv]

What does this PR do?

Type of change: ?
Replace the old calibration formula from t = a * exp(b * S) / L to t = 1 - exp(-a * (S / (1 - S))^b / seq_k^c. The old calibration breaks at short context. We can clamp at runtime or add 1 - exp(-...) wrapper around t because 1 − exp(−c/L^α) \in (0, 1)  strictly, so we don't need to hardcode where to cutoff.

Usage

Testing

The calibration curve are shown below.

Before your PR is "Ready for review"

Make sure you read and follow Contributor guidelines and your commits are signed (git commit -s -S).

Make sure you read and follow the Security Best Practices (e.g. avoiding hardcoded trust_remote_code=True, torch.load(..., weights_only=False), pickle, etc.).

• Is this change backward compatible?: ✅ / ❌ / N/A
• If you copied code from any other sources or added a new PIP dependency, did you follow guidance in CONTRIBUTING.md: ✅ / ❌ / N/A
• Did you write any new necessary tests?: ✅ / ❌ / N/A
• Did you update Changelog?: ✅ / ❌ / N/A
• Did you get Claude approval on this PR?: ✅ / ❌ / N/A

Additional Information

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 7e90e3b7-0d52-46cd-991a-70324a66a933

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch kaix/skip_softmax_calib_formula

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[github-actions]

PR Preview Action v1.8.1
🚀 View preview at https://NVIDIA.github.io/Model-Optimizer/pr-preview/pr-1541/
Built to branch gh-pages at 2026-05-25 01:50 UTC. Preview will be ready when the GitHub Pages deployment is complete.

[codecov]

Codecov Report

❌ Patch coverage is 84.26966% with 14 lines in your changes missing coverage. Please review.
✅ Project coverage is 76.80%. Comparing base (16a0130) to head (a42daf0).
⚠️ Report is 2 commits behind head on main.

Files with missing lines	Patch %	Lines
.../torch/sparsity/attention_sparsity/plugins/vllm.py	0.00%	5 Missing ⚠️
...kernels/sparsity/attention/ltx_triton_attention.py	50.00%	3 Missing ⚠️
.../attention_sparsity/methods/triton_skip_softmax.py	66.66%	3 Missing ⚠️
...s/sparsity/attention/diffusers_triton_attention.py	33.33%	2 Missing ⚠️
...rsity/attention_sparsity/calibration/calibrator.py	97.82%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1541      +/-   ##
==========================================
- Coverage   76.81%   76.80%   -0.02%     
==========================================
  Files         476      476              
  Lines       51891    51905      +14     
==========================================
+ Hits        39860    39864       +4     
- Misses      12031    12041      +10     

Flag	Coverage Δ
unit	52.71% <84.26%> (-0.01%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.