(seed, pid_v, pid_h, sample_idx, noise_offsets)
| 618 | ) |
| 619 | for peer_rank in tl.static_range(0, tp_world_size): |
| 620 | peer_base = tl.load(buffer_ptrs + peer_rank) |
| 621 | peer_maxs_ptr = peer_base.to(tl.pointer_type(tl.float32)) |
| 622 | peer_maxs_idx_ptr = peer_base.to(tl.pointer_type(tl.int64)) |
| 623 | tl.store( |
| 624 | peer_maxs_ptr + source_rank_base_offset, |
| 625 | gumbel_max, |
| 626 | mask=mask_h_out, |
| 627 | ) |
| 628 | tl.store( |
| 629 | peer_maxs_idx_ptr + storage_offset_maxs_idx + source_rank_base_offset, |
| 630 | gumbel_max_idx_global, |
| 631 | mask=mask_h_out, |
| 632 | ) |
| 633 | else: |
no outgoing calls
no test coverage detected