flashinfer.gated_act_mxfp8.silu_and_mul_mxfp8_quantize_backward

flashinfer.gated_act_mxfp8.silu_and_mul_mxfp8_quantize_backward(gated_input: Tensor, grad_output: Tensor, *, rowwise: bool = True, colwise: bool = False) → Tuple[Tensor, Tensor, Tensor, Tensor]

Apply the SwiGLU backward transform and emit RCEIL MXFP8 outputs.

The logical result concatenates the gate-input and up-input gradients, producing shape [M, 2K] before quantization. No intermediate BF16 result is materialized in global memory.

Parameters:

  • gated_input (torch.Tensor) – Contiguous BF16 CUDA tensor with shape [M, 2K].

  • grad_output (torch.Tensor) – Contiguous BF16 CUDA tensor with shape [M, K].

  • rowwise (bool, optional) – Whether to emit rowwise MXFP8 values and scales.

  • colwise (bool, optional) – Whether to emit colwise MXFP8 values and scales.

Returns:

Rowwise values, colwise values, rowwise scales, and colwise scales for the concatenated gate-input and up-input gradients. Outputs for disabled orientations are zero-sized tensors.

Return type:

tuple of torch.Tensor