nn.Conv2d: input [B, H, W, C_in] (NHWC), weight [C_out, kH, kW, C_in/groups] → [B, H', W', C_out]. mlx weight layout — the oracle transposes PyTorch's [C_out, C_in, kH, kW] at load. For the gemma-4 audio SSCP subsampler.
nn.Conv2d: input [B, H, W, C_in] (NHWC), weight [C_out, kH, kW, C_in/groups] → [B, H', W', C_out]. mlx weight layout — the oracle transposes PyTorch's [C_out, C_in, kH, kW] at load. For the gemma-4 audio SSCP subsampler.