o -i @sLddlZddlmZddlmZd ddZd dejd ejd ejfd d ZdS)N)_masked_compaction) Bitmatrixc Csd|j\}}t|}t|}t|tr|jj}t|f||||d|d||||d ||fS)un Return compacted copies of *yv* and *yi* based on a per-row bitmask. Only the elements whose index appears among the active bits of *bitmask* are kept; the rest are replaced by *sentinel*. Kept elements preserve their original left-to-right order. Parameters ---------- yv : torch.Tensor, shape (B, K) Values tensor. yi : torch.Tensor, shape (B, K), dtype torch.long Integer indices (0 ≤ index < 32) associated with *yv*. bitmask : torch.Tensor, shape (B,) **or** (B, 32) Per-row mask of active indices. See the in-place version for details. sentinel : int, default -1 Value written into dropped positions of the returned tensors. Returns ------- (yv_out, yi_out) : Tuple[torch.Tensor, torch.Tensor], each shape (B, K) New tensors with the same dtype/device as the inputs. rr)K) shapetorch empty_like isinstancerstoragedatarstride)yvyibitmasksentineln_rowsn_colsret_yvret_yirg/home/ubuntu/veenaModal/venv/lib/python3.10/site-packages/vllm/third_party/triton_kernels/compaction.py compactions    rrrrcCs|j\}}|j}dtjd||jd>}|d|@dk}|jdd} | d|} |  tj j ddd } |d| } |d| } | d| }|| |<|| |<| | fS) z6 reference implementation of `masked_compact` r )devicedtyperr) start_dimT)dimstable) rrraranger unsqueezeflattengatherlongtointargsort)rrrrBrrwbitsmaskkeeporder yi_sorted yv_sorted keep_sortedrrrcompaction_torch/s       r1)r)r%compaction_details._masked_compactionrtensorrrTensorr1rrrrs    ")