(
logits: T.Tensor((n_tok, n_exp), T.float32),
out_max: T.Tensor((n_tok,), T.float32),
out_sum: T.Tensor((n_tok,), T.float32),
)
| 56 | def gate_reduce(n_tok: int, n_exp: int): |
| 57 | @T.prim_func |
| 58 | def kernel( |
| 59 | logits: T.Tensor((n_tok, n_exp), T.float32), |
| 60 | out_max: T.Tensor((n_tok,), T.float32), |
| 61 | out_sum: T.Tensor((n_tok,), T.float32), |
| 62 | ) -> None: |
| 63 | with T.Kernel(n_tok, threads=32) as pid: |
| 64 | lf = T.alloc_fragment(n_exp, T.float32) |
| 65 | T.copy(logits[pid, 0], lf) |
| 66 | mx = T.alloc_fragment(1, T.float32) |
| 67 | T.reduce_max(lf, mx, dim=0) |
| 68 | sm = T.alloc_fragment(1, T.float32) |
| 69 | T.reduce_sum(lf, sm, dim=0) |
| 70 | if T.get_thread_binding() == 0: |
| 71 | out_max[pid] = mx[0] |
| 72 | out_sum[pid] = sm[0] |
| 73 | |
| 74 | return kernel |
| 75 |
no test coverage detected
searching dependent graphs…