CFK-38052: pin nvidia-cutlass-dsl to 4.5.0

[TorinLi]

Before submitting

• I agree to license this contribution under the terms of LICENSE.txt.
• I ran pre-commit run and committed any formatting changes.

Affected area

Build, packaging, or installation

Summary

Pin the optional CuTeDSL dependency to the known-good nvidia-cutlass-dsl[cu13]==4.5.0 release.

Why

The dependency was broadened from ==4.5.0 to >=4.5.0, which allows pip to resolve CUTLASS DSL 4.6.x. The libNVVM bundled with 4.6.x fails to compile the large SM100 grouped GEMM DGLU FP8 kernels, while the same kernels compile and pass with 4.5.0.

Restoring the exact pin prevents GPT OSS 20B and DeepSeek V3 environments from silently resolving to the affected compiler release while the forward compiler fix is being validated.

Related issues

Related to #397.

Internal tracking: NvBug 6503977 / CFK-38052.

API and compatibility impact

No public API changes. Installing the [cutedsl] extra now deterministically selects CUTLASS DSL 4.5.0 instead of accepting affected 4.6.x releases.

Testing

• git diff --check
• Verified the [project.optional-dependencies].cutedsl entry resolves to the exact known-good version.
• GPU kernel tests were not rerun for this dependency-only rollback; issue [Bug]: cutedsl kernel fails on upgrade to 4.6.0 #397 documents the same DGLU FP8 kernels passing with 4.5.0 and failing with 4.6.x.

Created by dkg-ai-bugfix agent.

Summary by CodeRabbit

• Chores
  • Pinned the optional NVIDIA CUTLASS DSL dependency to version 4.5.0 for more consistent installation behavior.

[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: 345bd359-d6e0-4d4c-b9d4-893a2e2588aa

📥 Commits

Reviewing files that changed from the base of the PR and between dd5ce86 and cdf0287.

📒 Files selected for processing (1)

• pyproject.toml

---

📝 Walkthrough

Walkthrough

The optional cutedsl dependency in pyproject.toml is changed from accepting version 4.5.0 or newer to requiring exactly version 4.5.0.

Changes

Dependency pin

Layer / File(s)	Summary
Pin optional CUTLASS DSL dependency pyproject.toml	The nvidia-cutlass-dsl[cu13] optional dependency is pinned from >=4.5.0 to ==4.5.0.

Estimated code review effort: 1 (Trivial) | ~2 minutes

Suggested reviewers: anerudhan

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately summarizes the main change: pinning nvidia-cutlass-dsl to 4.5.0.
Description check	✅ Passed	The description follows the template and includes the required sections with clear summary, rationale, related issue, impact, and testing.
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.}

[TorinLi]

@Anerudhan, could you please review this dependency rollback? It pins nvidia-cutlass-dsl to the known-good 4.5.0 release to contain the 4.6.x libNVVM regression tracked in #397. The change is intentionally limited to the dependency constraint; forward 4.5/4.6 API compatibility work remains tracked in #411. Thanks!

[Anerudhan]

The reason for bump is upstream FW like vLLM and SGLang have bumped to 4.6.x. Pinning the cuteDSL to 4.5.0 errors in the dependency solver.