b2b_causal_conv1d

b2b_causal_conv1d(x, weight_proj, weight_mixer, skip_bias)

Back-to-back fused causal 1D conv via the subq_ops CUDA kernel.

Fused kernel performing projection conv, pre-gate, mixer conv with skip, and post-gate. See upstream docstring for the exact algorithm.

Parameters:

• x (Tensor) – Input tensor [B, 3*C, L].

• weight_proj (Tensor) – Projection weight [3*C, K] (depthwise).

• weight_mixer (Tensor) – Mixer weight [C, K] (depthwise).

• skip_bias (Tensor) – Skip-bias scalar-per-channel [C].

Returns:

Output tensor [B, C, L].

Return type:

Tensor