| 121 | |
| 122 | @R.function |
| 123 | def entry_a( |
| 124 | q: R.Tensor((32, 8, 16, 8), dtype="float16"), |
| 125 | k: R.Tensor((32, 8, 16, 8), dtype="float16"), |
| 126 | v: R.Tensor((32, 8, 16, 8), dtype="float16"), |
| 127 | ) -> R.Tensor((32, 8, 16, 8), dtype="float16"): |
| 128 | cls = Expected |
| 129 | with R.dataflow(): |
| 130 | workspace_main: R.Tensor((65536,), dtype="uint8") = R.builtin.alloc_tensor( |
| 131 | R.shape([65536]), R.dtype("uint8"), R.prim_value(0) |
| 132 | ) |
| 133 | gv: R.Tensor((32, 8, 16, 8), dtype="float16") = cls.fused_relax_nn_attention_cutlass1( |
| 134 | q, k, v, workspace_main |
| 135 | ) |
| 136 | R.output(gv) |
| 137 | return gv |
| 138 | |
| 139 | @R.function |
| 140 | def entry_b( |