flashinfer.cake_gdn_det_train_kcp.cake_gdn_det_train_kcp_workspace_bytes¶
- flashinfer.cake_gdn_det_train_kcp.cake_gdn_det_train_kcp_workspace_bytes(ctx: Any, *, H_qk: int, HV: int, save_mode: str = 'light', boundary_mode: str = 'saved', save_transfer: bool = True, backward: bool = True, with_conv: bool = True, C: int | None = None, P: int | None = None, group_reduce: str = 'fused', reuse_recompute_tapes: bool | str | None = None, use_qk_l2norm_in_kernel: bool = False, fused_walks: int = 1, graph: str = 'off', save_for_backward: bool = True) int¶
Caller-held workspace bytes of one rank for
(ctx, H_qk, HV)(exchange pools, staging, tapes): the most the workspace binds in the forward or the backward of the given options (reuse_recompute_tapesas a keyword; the per-phase terms arecake_gdn_det_train_kcp_workspace_layoutof the package).use_qk_l2norm_in_kernel=Trueadds the transient normalized rowsqn/kn(2 x 256 T_r H_qkbytes) to the forward and every backward phase, and the8 T_r H_qk-byte FP32rwhere the backward re-runs the prep program.- Parameters:
ctx – This rank’s context from
cake_gdn_det_train_kcp_build_context().H_qk (int) – Number of query and key heads.
HV (int) – Number of value heads.
**kw – Backend options for this query, including the call’s save and boundary settings.