flashinfer.cake_kda_nondet_train.cake_kda_nondet_train_workspace_bytes¶
- flashinfer.cake_kda_nondet_train.cake_kda_nondet_train_workspace_bytes(S: int, num_docs: int, H: int, *, backward: bool = True, save_mode: str = 'light', sm_count: int | None = None, arena: Any | None = None, doc_lengths: Sequence[int] | None = None, use_gate_in_kernel: bool = False, use_qk_l2norm_in_kernel: bool = False) int¶
Bytes
cake_kda_nondet_train_workspace()allocates for the problem (sm_count: the device’s SM count, the current CUDA device’s whenNone; it sets the segment count of the segmented recurrences;arena: the lower-memory arena option, off whenNone;doc_lengths: the exact document groups of a batched option, else their bound;use_gate_in_kernel: the gate mode, whose backward layouts carry no gate epilogue buffers in raw mode;use_qk_l2norm_in_kernel: the Q/K L2 norm mode, likewise without those buffers).