nn.Conv3d: input [B, D, H, W, C_in] (NDHWC), weight [C_out, kD, kH, kW, C_in/groups] → [B, D', H', W', C_out]. The Qwen3-VL vision patch embed (kernel == stride) — the conv KERNEL's accumulation order is the parity- relevant arithmetic; a value-equivalent GEMM differs by bf16 ulp.
nn.Conv3d: input [B, D, H, W, C_in] (NDHWC), weight [C_out, kD, kH, kW, C_in/groups] → [B, D', H', W', C_out]. The Qwen3-VL vision patch embed (kernel == stride) — the conv KERNEL's accumulation order is the parity- relevant arithmetic; a value-equivalent GEMM differs by bf16 ulp.