Method forward

cake-core/src/utils/fp8.rs:232–243

Forward pass: dequantizes weight if needed, computes matmul in weight dtype. The cached weight is already transposed and contiguous for optimal matmul.

(&self, x: &candle_core::Tensor)

    /// Forward pass: dequantizes weight if needed, computes matmul in weight dtype.
    /// The cached weight is already transposed and contiguous for optimal matmul.
    pub fn forward(&self, x: &candle_core::Tensor) -> candle_core::Result<candle_core::Tensor> {
        let in_dtype = x.dtype();
        let wt = self.get_weight()?; // pre-transposed
        let compute = wt.dtype();
        let x = if x.dtype() == compute { x.clone() } else { x.to_dtype(compute)? };
        let y = x.broadcast_matmul(&wt)?;
        let y = match &self.bias {
            Some(b) => y.broadcast_add(&b.to_dtype(compute)?)?,
            None => y,
        };
        if in_dtype == compute { Ok(y) } else { y.to_dtype(in_dtype) }
    }
}

/// Create an Fp8Linear from a VarBuilder (no bias).