Function silu_mul_quant_fp8 

pub fn silu_mul_quant_fp8(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    gate: DevicePtr,
    up: DevicePtr,
    out_fp8: DevicePtr,
    a_scale: DevicePtr,
    out_bf16: DevicePtr,
    m: u32,
    k: u32,
    stream: u64,
) -> Result<()>

Fused SiLU·mul + per-token-group(128) FP8-E4M3 quantization — replaces the silu_mul → per_token_group_quant_fp8 pair on the W8A8 prefill down-path without materializing the BF16 intermediate. Bit-identical to the pair (product rounds through BF16 before the group max; same reduction order, scale floor, and SATFINITE encode).

out_bf16 is nullable (DevicePtr::NULL): pass the post-SiLU BF16 buffer only when a downstream consumer needs it (expert down_proj LoRA fold).

Kernel: silu_mul_quant_fp8(gate, up, out_fp8, a_scale, out_bf16, M, K) Grid: (M, 1, 1) Block: (128, 1, 1). Caller must ensure k % 128 == 0 and k / 128 <= 16 (SILU_QUANT_MAX_GROUPS) — fall back to the unfused pair otherwise.