flashinfer.cake_dsa_sparse_attention_nondet_train.cake_dsa_sparse_attention_nondet_nope_varlen

flashinfer.cake_dsa_sparse_attention_nondet_train.cake_dsa_sparse_attention_nondet_nope_varlen(q: Tensor, kv: Tensor, gather_kv_indices: Tensor, cu_seqlens_q: Tensor, cu_seqlens_k: Tensor, max_seqlen_q: int | None = None, max_seqlen_k: int | None = None, *, causal: bool = False, topk_length: Tensor | None = None, softmax_scale: float | None = None, return_lse: bool = False, key_passes: int | None = None, backend: str = 'cake', dkv_acc: Tensor | Callable | None = None, dkv_dst_map: Tensor | None = None, dkv_direct: bool | None = None, check_dst_map: bool | None = None, check_cu_seqlens: bool = False)

Packed multi-document form of cake_dsa_sparse_attention_nondet_nope() (SM100 / SM103 / SM107).

Documents are packed along the first dimension of q (T = cu_seqlens_q[-1]) and of kv (S = cu_seqlens_k[-1]; [S, 512] or [S, 1, 512]); cu_seqlens_q and cu_seqlens_k are independent and query and key lengths may differ per document: the query segment is the tail of its key prefix, so query local_q of document d sits at key position (seqlen_k[d] - seqlen_q[d]) + local_q. gather_kv_indices [T, K_slots] int32 hold key positions relative to the row’s document (-1 or >= seqlen_k[d] invalid; any K_slots >= 1) and are offset by cu_seqlens_k on device. The route follows the table size (cake_backend.NOPE_CONVERT_SLOT_LIMIT slots): below the limit, without caller lengths, the forward kernel converts the rows itself in its single launch, position-preserving (the global key or -1 at the slot it came from; the row’s length is its last valid slot + 1); at or above it, and always with a caller topk_length, one launch right before the forward kernel compacts each row (valid keys first, in slot order, then -1) so rows with holes cost their valid slots, and the caller’s topk_length is a position cut applied before compaction. Zero-length query segments contribute no rows; a row whose every slot is invalid gives out = 0, lse = -inf and zero gradients. The kernels apply no positional mask of their own; the offset index rows define the key sets. max_seqlen_q / max_seqlen_k are accepted for signature parity with other varlen attention entry points and are not read on the host. See cake_dsa_sparse_attention_nondet_nope() for the other arguments, the 1 / 16 default scale, the FP32 dK + dV semantics and the returns.

Under torch.use_deterministic_algorithms(True) the forward runs and the backward raises PyTorch’s RuntimeError before its first backward kernel launch (its dK / dV atomics are not run-to-run reproducible); deterministic preparation may run first (see the module docstring). With warn_only=True it issues a UserWarning and runs. No deterministic package is chosen instead.

cu_seqlens_q and cu_seqlens_k are caller-owned prefix sums with these invariants: int32 [num_docs + 1] tensors of one length on the device of q, starting at 0 and non-decreasing, cu_seqlens_q[-1] == T (the packed rows of q), cu_seqlens_k[-1] == S (the packed rows of kv; a shorter or longer key window is not accepted) and seqlen_q <= seqlen_k for every document. Their structure (contiguous int32 [num_docs + 1] of one length) is checked on every call; their values only with check_cu_seqlens=True, since that check synchronizes with the device, and a violation is otherwise not detected on device: a document with seqlen_k < seqlen_q gives negative own positions (all-invalid rows under causal), a non-monotonic prefix sum assigns rows to the wrong document. The per-document rows are turned into global key rows on device: by the forward kernel itself below the slot limit (position-preserving, one launch), else compacted in one launch right before the forward kernel; a caller topk_length is a position cut applied before compaction and the kernels then read the compacted rows and their counts.

Parameters:
  • check_cu_seqlens (bool) – True checks the values of cu_seqlens_q / cu_seqlens_k against the invariants above at the cost of one blocking device-to-host copy per call, before any launch (skipped during CUDA-graph capture, where no synchronization is possible; the eager call captured was checked). False (default) keeps the structural checks (dtype, rank, length, contiguity, device) only.

  • causal (bool) – With False (default) the index row is taken as is – only -1 / out-of-range slots are dropped. With True a slot that selects a key after the query’s own position (idx > (seqlen_k[d] - seqlen_q[d]) + local_q) is invalid too, so a top-k selector need not enforce causality itself.

  • dkv_direct (Optional[bool]) – As in cake_dsa_sparse_attention_nondet_nope(): None (default) reads FLASHINFER_CAKE_DSA_DKV_DIRECT; True / False decide explicitly.

  • check_dst_map (Optional[bool]) – As in cake_dsa_sparse_attention_nondet_nope(): None (default) reads FLASHINFER_CAKE_DSA_CHECK_DST_MAP; True validates the values of dkv_dst_map on every backward, False never does.