SubqOpsCausalConv1d#

class SubqOpsCausalConv1d(
in_channels,
out_channels,
kernel_size,
*,
groups,
bias=False,
activation='identity',
)#

Bases: Conv1d

Depthwise causal 1D conv using subquadratic_ops_torch.causal_conv1d.

Parameters:
__init__(
in_channels,
out_channels,
kernel_size,
*,
groups,
bias=False,
activation='identity',
)#

Build a depthwise causal conv with optional SiLU activation.

Parameters:

  • in_channels (int) – Channel count (must equal out_channels and groups).

  • out_channels (int) – Output channels (depthwise: same as in_channels).

  • kernel_size (int) – Causal kernel length.

  • groups (int) – Must equal in_channels for depthwise layout.

  • bias (bool) – Whether to include a per-channel bias.

  • activation (str) – "identity" or "silu" applied inside the CUDA kernel.

Return type:

None

forward(input)#

Run causal depthwise conv; input shape [B, C, L].

Casts the weight and bias to input’s dtype when they differ. The fused kernel selects its specialisation from the INPUT dtype and then requires every tensor to match it exactly, so under torch.autocast — where activations arrive as bf16/fp16 while parameters stay fp32 — passing the parameters through unchanged raises:

ValueError: in_w expected dtype (code=4, bits=16) but got (code=2, bits=32)

Casting here mirrors what autocast does for the built-in conv ops: the fp32 master parameters are untouched, only the values handed to the kernel are narrowed.

Parameters:

input (Tensor)

Return type:

Tensor