pub fn ffn_act_keep(
model: &Arc<CmfModel>,
w1: usize,
w3: usize,
w2: usize,
xs: &[f32],
b: usize,
hidden: usize,
inter: usize,
act: GraphAct,
out: &mut [f32],
) -> boolExpand description
The LLM prefill’s dense FFN for separate gate/up/down weights in any codec with a device GEMM (int8, q4tp, or a mix of them): the panels stay on the card between the projections. wgpu only; other backends decline and the caller keeps its per-GEMM path.