[YAMY1234]

Motivation

The CuTe DSL BF16 GEMM backend is selected on SM10x GPUs, but both runtime entry points reject SM107. Enabling the architecture gate also exposes a bias conversion issue: DLPack rejects a bias view that retains requires_grad=True, even during inference under torch.no_grad().

Modifications

• Allow SM107 in both CuTe DSL BF16 GEMM runtime entry points.
• Detach the read-only bias view before exporting it through DLPack. This does not add a copy or synchronization.
• Update backend diagnostics and CLI help to describe SM10x support.
• Extend the existing kernel test with an SM107 workload shape and a requires_grad=True bias under torch.no_grad().

The GEMM kernel math, tactic selection, and launch behavior are unchanged.

Accuracy Tests

The existing numerical comparison test now covers the added shape and gradient-bearing bias case. Controlled SM107 validation passed for eager execution, CUDA Graph execution, and an end-to-end accuracy workload on the pinned development base.

Current upstream-main GPU CI is pending while this PR remains a draft.

Speed Tests and Profiling

Controlled end-to-end SM107 testing confirmed that enabling this backend removes the fallback to torch.nn.functional.linear. No kernel implementation or tactic was changed.

Current upstream-main performance confirmation is pending while this PR remains a draft.

Checklist

• Format the code with pre-commit.
• Extend the existing unit test.
• Update user-facing backend documentation.
• Complete GPU CI on current upstream main.
• Add current-main performance confirmation.

Local Validation

• pre-commit run --files passed for all four modified files.
• git diff --check passed.
• Python compilation passed for all four modified files.

---

CI States

Latest PR Test (Base): ❌ Run #31033970223
Latest PR Test (Extra): ❌ Run #31033969425

[b8zhong]

Can we just change is to is_sm100_supported?

[leejnau]

maybe is_sm10x_supported ? what is Supports 110 exactly?

[mmangkad]

is_sm10x_supported would’ve been nice if there were no need to check different cuda versions, but since sm107 needs >= 13.4 while is_sm100_supported currently checks >= 12.8 (actually, sm103 needs >= 12.9), maybe for clarity rename it to is_sm100_or_sm103_supported and add is_sm107_supported

[b8zhong]

@leejnau You can ignore the other part of my comment. Was typing something else on another tab or something..

[b8zhong]

@mmangkad I don't think we really need to check the CUDA version. It should be correct already (@leejnau What do you think?)

[YAMY1234]

Thanks! I updated the PR to use is_sm100_supported() instead of explicitly checking for SM100/103/107.

BTW, Lee and I talked about it, and we don’t think we need a separate is_sm107_supported() for now. If we add features that are only SM107-specific later, we can add it then as a follow up

[mmangkad]

Yeah maybe not for now, but at some point we'll probably have to distinguish them

[leejnau]

I tested Kimi K3 on sm107 and the cutedsl bf16 gemm works ✅

[YAMY1234]

/tag-and-rerun-ci