Function silu_mul 

pub fn silu_mul(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    gate: DevicePtr,
    up: DevicePtr,
    output: DevicePtr,
    num_elements: u32,
    stream: u64,
) -> Result<()>

2026-09-25: output = act(gate) * up over num_elements BF16 values, where kernel fixes the activation: callers pass moe_silu_mul (SiLU) or, for a GeLU FFN, gelu_mul. The moe_silu_mul.cu copies in gb10/deepseek-v4-flash (also compiled for longcat-flash-lite) and gb10/step3p7-flash first clamp gate to at most 10 and up to [-10, 10].