Add split-k decode kernel for skip softmax

[kaix-nv]

What does this PR do?

Type of change: ?
New feature.

Adds skip-softmax support to the dedicated paged decode kernel and routes compatible one-token vLLM decode launches
through it.

• Uses one sequential KV split with 32-token tiles so skip decisions follow the running softmax maximum.
• Preserves the existing 32-split, 128-token schedule when skip-softmax is disabled.
• Supports skip-softmax together with NVFP4 P/V QDQ.
• Validates thresholds and passes them as runtime kernel arguments to avoid threshold-dependent recompilation.
• Keeps prefill, multi-token decode, and unknown sparse options on the shared Triton attention kernel.
• Preserves the existing FlashAttention and FlashInfer fallback behavior.

Usage

 from modelopt.torch.kernels.common.attention.decode_attention import attention_decode

  output = attention_decode(
      query,
      key_cache,
      value_cache,
      block_table,
      sequence_lengths,
      skip_softmax_threshold=0.1,
      p_qdq="nvfp4",
      p_qdq_amax=1.0,
      v_qdq="nvfp4",
      v_qdq_amax=1.0,
  )

Testing

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: 55ed7de4-8ac1-4b17-8ced-62e028eb4f4c

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/splitk_skip_softmax

---

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

[github-actions]

PR Preview Action v1.8.1
🚀 View preview at https://NVIDIA.github.io/Model-Optimizer/pr-preview/pr-1941/
Built to branch gh-pages at 2026-07-08 01:30 UTC. Preview will be ready when the GitHub Pages deployment is complete.

[codecov]

Codecov Report

❌ Patch coverage is 3.92157% with 441 lines in your changes missing coverage. Please review.
✅ Project coverage is 60.59%. Comparing base (75b5803) to head (117725f).
⚠️ Report is 11 commits behind head on main.

Files with missing lines	Patch %	Lines
.../torch/sparsity/attention_sparsity/plugins/vllm.py	0.00%	206 Missing ⚠️
...torch/kernels/common/attention/decode_attention.py	0.00%	134 Missing ⚠️
...lopt/torch/kernels/quantization/attention/v_qdq.py	0.00%	40 Missing ⚠️
...delopt/torch/kernels/common/attention/triton_fa.py	31.57%	39 Missing ⚠️
modelopt/torch/quantization/plugins/vllm.py	0.00%	18 Missing ⚠️
...t/torch/kernels/quantization/common/nvfp4_quant.py	0.00%	4 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1941      +/-   ##
==========================================
- Coverage   61.21%   60.59%   -0.62%     
==========================================
  Files         515      521       +6     
  Lines       57245    61650    +4405     
==========================================
+ Hits        35043    37359    +2316     
- Misses      22202    24291    +2089     

Flag	Coverage Δ
unit	55.02% <3.92%> (+0.10%)	⬆️

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

☔ View full report in Codecov by Harness.
📢 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.