CSA: request ex2.approx.ftz through fastmath= so the compressor builds at the cutlass-dsl floor

[zkyue]

cute.math.exp2's approx/ftz keywords are newer than the
nvidia-cutlass-dsl[cu13]>=4.5.0 floor python/pyproject.toml resolves to, so the
ratio=128 compressor's fast-exp path raised TypeError: exp2() got an unexpected keyword argument 'approx' on any environment that installs that floor -- 14 failures
in test/python/fe_api/csa/test_CSA_compressor.py, all from this one call site (#460).

fastmath=True is accepted by both versions and lowers to the same
ex2.approx.ftz.f32 the r128 tolerance contract is calibrated on, so this is a pure
spelling change:

• cutlass-dsl 4.5.0 before: 14 failed, 74 passed. After: 88 passed, 1 skipped.
• cutlass-dsl 4.6.1 after: the 16-kernel reg_probe_csa_compressor_r128.py PTX is
  byte-identical to before this change -- same register counts, 0 spill / 0 stack, and
  1104 ex2.approx.ftz.f32 across the set -- so nothing changes on 4.6.
• cutlass-dsl 4.5.0 after: the same 1104 ex2.approx.ftz.f32, and no no-.ftz
  ex2.approx.f32 anywhere.
• gate_csa_compressor_r128.py: GATE PASS 21/21 on both 4.5.0 and 4.6.1.

Verified on B200 (CC 10.0), cuDNN 9.23, driver 590.48.01.

Signed-off-by: zky kaiyue.zhou@z.ai

This addresses the CSA half of #460 only (the 14 test_CSA_compressor.py failures attributed to b950af1). It does not touch the 25 wgrad failures from #456 -- those reproduce on 4.6.1 as well per the table in the issue, and cutlass.utils.rubin_helpers is a separate gap.

Summary by CodeRabbit

• Bug Fixes
  • Improved compatibility of the fast exponential calculation while preserving approximate, flush-to-zero behavior.
  • No user-facing API changes.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 59978d4e-a85f-42dd-82e5-248c3f444fa9

📥 Commits

Reviewing files that changed from the base of the PR and between 968f1ef and d50f12e.

📒 Files selected for processing (1)

• python/cudnn/csa/compressor/compressor_sm100_r128.py

---

📝 Walkthrough

Walkthrough

The _exp_fast function in compressor_sm100_r128.py now calls cute.math.exp2 with the fastmath=True keyword instead of approx=True, ftz=True. The compatibility comments are updated accordingly. The FMA-based exponent transformation logic remains unchanged.

Changes

exp2 invocation change

Layer / File(s)	Summary
Update exp2 fastmath keyword python/cudnn/csa/compressor/compressor_sm100_r128.py	Changes the exp2 call in _exp_fast from approx=True, ftz=True to fastmath=True and updates related comments.

Estimated code review effort: 1 (Trivial) | ~3 minutes

Suggested reviewers: anerudhan

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly identifies the CSA compressor change and the compatibility goal for the cutlass-dsl dependency floor.
Description check	✅ Passed	The description explains the problem, solution, compatibility impact, issue scope, and detailed test results, despite not using the template headings.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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