[Bug]: CuTe DSL NVFP4 SwiGLU / FP4Out dense GEMM has no SM107 kernel

[farazkh80]

Summary

The CuTe DSL NVFP4 dense GEMM with fused SwiGLU (and the FP4-output variant) is Blackwell-only. On SM107 the ops raise at init:

CuteDSL NVFP4 SwiGLU backend requires SM 100 (B200) or SM 103 (B300), but got SM 107
CuteDSL NVFP4 SwiGLU FP4Out requires SM 100 or SM 103

Gates: tensorrt_llm/_torch/custom_ops/cute_dsl_custom_ops.py around L1716 and L2248.

Impact

TestDeepSeekV3Lite::test_cute_dsl_nvfp4[...] cannot run on SM107 (found while sweeping llm_function_core.txt on 4x SM107 against main). The dense BF16/FP8/MXFP8/NVFP4 SM107 GEMMs landed in #18761/#18765, but the SwiGLU act-fusion variants did not.

Ask

• Port the NVFP4 SwiGLU / FP4Out act-fusion dense GEMM to SM107 (same family as [None][feat] Add SM107 CuTe DSL quantized dense GEMM/BMM custom ops and dispatch #18765), or
• keep the SM100/103 gate and make the skip reason in test_llm_api_pytorch.py (~L1116, currently "blockscaling mm supports SM 100 and 103 only") say that the SwiGLU fusion is the missing piece.

[100milliongold]

I took the second of your two options and opened #19448 for it: the skip reason in both test_cute_dsl_nvfp4 sites now names the act fusion rather than blockscaling matmul as a whole.

-            pytest.skip("CuTe DSL blockscaling mm supports SM 100 and 103 only")
+            pytest.skip(
+                "CuTe DSL NVFP4 GEMM+SwiGLU act fusion supports SM 100 and 103 only"
+            )

Two sites, no behaviour change. I left the message self-describing instead of linking this issue, because the skips in that file reference https://nvbugs/... rather than GitHub issue numbers.

I did not touch the gates. Comparing what #18761 and #18765 actually did, they added SM107-specific ops and dispatch with their own *_rubin test files rather than widening the existing (100, 103) checks, so the act-fusion port is the same shape of work and belongs with the owners of those kernels. Flipping cute_dsl_custom_ops.py to is_sm_100f() would only move the failure past the gate.

So this leaves your first option open. @peaceh-nv, if the port is planned, feel free to close #19448 and let the accurate message come with it instead.