[GG] build: restore CUTLASS DSL 4.6.0 after B12X migration

[voipmonitor]

Summary

• Revert [GG] build: pin CUTLASS DSL 4.5.3 for SM120 W4A16 #128's CUTLASS DSL 4.5.3 pin.
• Restore nvidia-cutlass-dsl[cu13]==4.6.0 on dev/gilded-gnosis.
• Remove the now-stale spill warning added by [GG] build: pin CUTLASS DSL 4.5.3 for SM120 W4A16 #128.

Why

#128 was correct for the pre-migration B12X W4A16 kernel: CUTLASS DSL 4.6.0 produced a 255-register kernel with a 256-byte per-thread stack and reduced TP8 A16 64k prefill throughput to about 5,052 tok/s.

Current B12X master contains the explicit CUTLASS DSL 4.6 migration (6627d34). The tested B12X release commit bd494bc contains current master 957b9fd; its additional commits do not touch W4A16 or CUTLASS dependencies.

Validation

Exact stack:

• vLLM: 30680faf5cc1fa3122b08749f8a35eab19f63576
• B12X: bd494bce8eb7505d0955e8139ecc200093017d9f
• CUTLASS DSL: 4.6.0
• GLM-5.2 Luke NVFP4, TP8/DCP1/MTP0, force-A16
• clean JIT cache

64k standalone prefill after warmup:

Run	tok/s
1	5,916
2	5,906
3	5,904
Median	5,906

The previous 4.5.3 median was 5,909 tok/s, so the difference is measurement noise. The old regressed 4.6.0 result was 5,052 tok/s.

All 40 compiled W4A16FusedMoeKernel objects reported STACK:0 LOCAL:0. Register counts were 124, 128, 184, 220, or 247; the old REG:255 STACK:256 artifact was not reproduced.

Checks

• git diff --check
• exact one-file revert verified
• clean-cache TP8 A16 E2E benchmark
• direct cuobjdump --dump-resource-usage inspection of all generated W4A16 objects

Summary by CodeRabbit

• Chores
  • Updated the CUDA tooling dependency to version 4.6.0.
  • Removed outdated version-specific dependency notes.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: e1c027a5-3dcd-4379-a460-614283c9c7da

📥 Commits

Reviewing files that changed from the base of the PR and between 2217d94 and e434a9f.

📒 Files selected for processing (1)

• requirements/cuda.txt

---

📝 Walkthrough

Walkthrough

Updates the CUDA requirements pin for nvidia-cutlass-dsl[cu13] from 4.5.3 to 4.6.0 and removes the version-specific explanatory comment.

Changes

CUDA dependency update

Layer / File(s)	Summary
Update CUDA CUTLASS DSL pin requirements/cuda.txt	The pinned nvidia-cutlass-dsl[cu13] version changes to 4.6.0, and the adjacent version rationale comment is removed.

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

Possibly related PRs

• local-inference-lab/vllm#128: Updates the same CUDA dependency pin between versions 4.5.3 and 4.6.0.

Suggested reviewers: brandonpelfrey, jinzhen-lin, lukealonso, arpera, ameenp

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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