Remove --skip-softmax option from hf_sa (only allow calibration option)

[rohansjoshi]

What does this PR do?

The hf_sa.py script applies an attention sparsity method to a model and exports a "sparsified" checkpoint. The SKIP_SOFTMAX_DEFAULT sparse attention config doesn't run calibration or make any change to the checkpoint. If we use this method with hf_sa.py the exported checkpoint is identical to the original checkpoint. This could cause confusion so it's best to remove the option altogether. This PR removes the option and switches to SKIP_SOFTMAX_CALIB as the default config.

Summary by CodeRabbit

• Chores
  • Updated sparse attention configuration in example script: removed "skip_softmax" option and changed default to "skip_softmax_calib". The script no longer accepts the previous "skip_softmax" configuration path.

[coderabbitai]

📝 Walkthrough

Walkthrough

The change removes deprecated "skip_softmax" sparse attention configuration support from the HuggingFace sparse attention example script. The --sparse_attn CLI argument default is updated from "skip_softmax" to "skip_softmax_calib", and associated imports and configuration choices are adjusted accordingly.

Changes

Cohort / File(s)	Summary
Sparse Attention Configuration Cleanup examples/llm_sparsity/attention_sparsity/hf_sa.py	Removed SKIP_SOFTMAX_DEFAULT import, eliminated "skip_softmax" from SPARSE_ATTN_CFG_CHOICES, and updated --sparse_attn CLI default from "skip_softmax" to "skip_softmax_calib".

Estimated code review effort

🎯 1 (Trivial) | ⏱️ ~3 minutes

🚥 Pre-merge checks | ✅ 4

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately describes the main change: removing the --skip-softmax option and retaining only the calibration option in hf_sa.py.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Security Anti-Patterns	✅ Passed	Pull request introduces only configuration changes to hf_sa.py with no security anti-patterns or new dependencies.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch rohjoshi/rm-skipsoftmax-option

---

_{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-1123/
Built to branch gh-pages at 2026-03-25 20:32 UTC. Preview will be ready when the GitHub Pages deployment is complete.

[coderabbitai]

🧹 Nitpick comments (1)

examples/llm_sparsity/attention_sparsity/hf_sa.py (1)

230-235: Clarify CLI migration in help text.

Line 232 now defaults to a single supported value, which is fine. Consider explicitly stating that skip_softmax was removed to make CLI errors/migration clearer for existing users.

Suggested tweak

     parser.add_argument(
         "--sparse_attn",
         type=str,
         default="skip_softmax_calib",
         choices=list(SPARSE_ATTN_CFG_CHOICES.keys()),
-        help="Sparse attention configuration to apply.",
+        help="Sparse attention configuration to apply. Only 'skip_softmax_calib' is supported (legacy 'skip_softmax' removed).",
     )

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@examples/llm_sparsity/attention_sparsity/hf_sa.py` around lines 230 - 235,
Update the CLI help for the "--sparse_attn" argument to note the migration from
the removed "skip_softmax" option to the new default "skip_softmax_calib" so
users get a clear migration message; modify the help string where the argument
is defined (the "--sparse_attn" parser entry that uses SPARSE_ATTN_CFG_CHOICES
and default="skip_softmax_calib") to explicitly state that "skip_softmax" was
removed and that "skip_softmax_calib" is now the default, and keep the choices
list unchanged to avoid surprises.

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Nitpick comments:
In `@examples/llm_sparsity/attention_sparsity/hf_sa.py`:
- Around line 230-235: Update the CLI help for the "--sparse_attn" argument to
note the migration from the removed "skip_softmax" option to the new default
"skip_softmax_calib" so users get a clear migration message; modify the help
string where the argument is defined (the "--sparse_attn" parser entry that uses
SPARSE_ATTN_CFG_CHOICES and default="skip_softmax_calib") to explicitly state
that "skip_softmax" was removed and that "skip_softmax_calib" is now the
default, and keep the choices list unchanged to avoid surprises.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 68d1cedc-9562-4e37-b7a8-6f0c82c8a5f9

📥 Commits

Reviewing files that changed from the base of the PR and between 291498b and 03b8517.

📒 Files selected for processing (1)

• examples/llm_sparsity/attention_sparsity/hf_sa.py

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 70.18%. Comparing base (291498b) to head (03b8517).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1123      +/-   ##
==========================================
- Coverage   70.21%   70.18%   -0.03%     
==========================================
  Files         228      228              
  Lines       25952    25952              
==========================================
- Hits        18221    18215       -6     
- Misses       7731     7737       +6     

☔ 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.