[FEA] [DSL] Request to support sm_121

[HaoKang-Timmy]

Which component requires the feature?

CuTe DSL

Feature Request

Hi, I wonder if Cutedsl plans to support Nvidia DGX Spark. The newest hardware provided by Nvidia.

[mnicely]

We will absolutely support Spark. Let me confirm ETA and circle back.

[github-actions]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.

[github-actions]

This issue has been labeled inactive-90d due to no recent activity in the past 90 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed.

[Teora]

+1 Still urgently needed in 2026!

Multiple users are now reporting this issue as GB10 (DGX Spark) systems are being deployed:

• [BUG] cutlass.cute.nvgpu.common.OpError: OpError: expects arch to be one of ['sm_100a', 'sm_100f'], but got sm_121a #2802 - Multiple tcgen05 operations reject sm_121
• [BUG] [Python DSL] BlockScaledMmaOp restricts FP4 operations to sm_100a only, blocks sm_120/sm_121 #2800 - BlockScaledMmaOp blocks FP4 on sm_121

Hardware reality: GB10 supports FP4 Tensor Cores (1 PFLOPS @ FP4), but CuTe DSL currently blocks access to these capabilities.

Community impact: Growing number of DGX Spark users unable to use FP4 features.

Could the NVIDIA team provide an update on the timeline for sm_121 support? This seems like a straightforward addition since:

1. Arch.sm_121 is already defined in cutlass/base_dsl/arch.py
2. The hardware supports all the operations
3. Only admissible_archs lists need updating

Thank you for considering this feature request!

[swtb3-ryder]

@mnicely were you able to learn any information RE timelines?

If Ive understood community findings, the bottleneck for NVFP4 working on spark at full speed is in the support here. There hasn't really been much in terms of an update or timeline from NVIDIA regarding proper spark support. Which I am sure is quite frustrating for many spark users.