fix(sgl-kernel): use >= 120 for SM12x CUDA kernel dispatch (#18750)

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
2026-02-16 00:44:47 +08:00
parent 5fc328465a
commit 0d30896015
3 changed files with 3 additions and 3 deletions
@@ -665,7 +665,7 @@ void cutlass_fp4_group_mm(
          N,
          K);
    }
-  } else if (sm_version == 120) {
+  } else if (sm_version >= 120) {
    if (output.scalar_type() == torch::kBFloat16) {
      run_fp4_blockwise_scaled_group_mm_sm120(
          output,