Method warmup

cake-core/src/utils/fp8.rs:223–228

Pre-dequantize F8→F16 and cache. Call once before inference loop.

(&mut self)

    /// Pre-dequantize F8→F16 and cache. Call once before inference loop.
    pub fn warmup(&mut self) -> candle_core::Result<()> {
        let w = self.get_weight()?;
        *self.weight.write().unwrap() = Some(w);
        self.f8_weight = None; // Free the F8 storage
        Ok(())
    }

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