[Bug][v0.6.17][gb300]tests/gemm/test_groupwise_scaled_gemm_fp8.py:201: Mismatched elements: 124 / 8192 (1.5%)

### Summary Found this issue in FlashInfer CI. ### CI metadata Test case: tests/gemm/test_groupwise_scaled_gemm_fp8.py::test_fp8_groupwise_gemm_small_batch_size[K-256-256-32] FlashInfer commit: d4cd4c764bdaf8d59bb9a7c2ccbdf0bf4d36a8c6 Pipeline: [pipeline](https://nv/flashinfer-ci/-/pipelines/61491878) Job name(s): unit_test_gb300: [cu130] Branch: release-v0.6.17 Environment: gb300 ### Failed Jobs [unit_test_gb300](https://nv/flashinfer-ci/-/jobs/388035432): gb300 / cu130 ### Failure ```shell Mismatched elements: 124 / 8192 (1.5%) Greatest absolute difference: 1.71875 at index (8, 42) (up to 0.01 allowed) Greatest relative difference: 90.0 at index (8, 123) (up to 0.01 allowed) /workspace/flashinfer/tests/gemm/test_groupwise_scaled_gemm_fp8.py:201: AssertionError: Tensor-likes are not close! ``` ### Reproduction command ```python pytest 'tests.gemm.test_groupwise_scaled_gemm_fp8' ```