flashinfer.cake_causal_conv1d_nondet_train.cake_causal_conv1d_nondet_train_backward¶
- flashinfer.cake_causal_conv1d_nondet_train.cake_causal_conv1d_nondet_train_backward(x: Tensor, weight: Tensor, cu_seqlens: Tensor, dy: Tensor, *, bias: Tensor | None = None, silu_grad_preactivation: str = 'bf16', workspace: Tensor | None = None, dx: Tensor | None = None, dweight: Tensor | None = None, dbias: Tensor | None = None, bwd_grid_mult: int = 1) tuple[Tensor, ...]¶
Packed width-4 depthwise causal convolution + SiLU, training backward (one launch, unordered
dweight/dbiasreductions).Under
torch.use_deterministic_algorithms(True)the backward raises PyTorch’sRuntimeErrorbefore it launches (aUserWarning, then it runs, withwarn_only=True); it never falls back to the deterministic package.- Parameters:
x – The forward’s inputs; nothing else is saved.
weight – The forward’s inputs; nothing else is saved.
cu_seqlens – The forward’s inputs; nothing else is saved.
bias – The forward’s inputs; nothing else is saved.
dy (torch.Tensor) – BF16 gradient of
y,x’s shape, contiguous.silu_grad_preactivation ({"bf16", "fp32"}) –
"bf16": SiLU’ at the BF16-rounded recomputed preactivation (FP32 accumulators)."fp32": at the unrounded FP32 preactivation with FP64 accumulators.workspace (torch.Tensor, optional) – Caller-owned buffer from
cake_causal_conv1d_nondet_train_backward_workspace().dx (torch.Tensor, optional) – Preallocated BF16 output of
x’s shape.dweight (torch.Tensor, optional) – Preallocated
[C, 4]output (FP32 or BF16; the parameter’s dtype when omitted). An FP32dweightmust be 16-byte aligned (one 16-byte store per channel;ValueErrorotherwise).dbias (torch.Tensor, optional) – Preallocated
[C]output (FP32 or BF16; the parameter’s dtype when omitted).bwd_grid_mult ({1, 2, 3, 4}) – Opt-in finer static partition of the token axis (default 1, the default grid): the backward launches
bwd_grid_multtimes its default untrimmed CTA count, so each CTA streams that much fewer tokens (for a backward that overlaps kernels of other streams: a CTA that starts late trails by a shorter stream).dxis bitwise the default’s; more CTAs add into the same accumulators (same budgets, same workspace size).TypeErrorfor a non-int,ValueErrorfor another value.
- Returns:
(dx, dweight), with a third elementdbiasonly whenbiasis given.S == 0returns the emptydxand zero gradients without launching.- Return type:
tuple[torch.Tensor, …]