Method dispatch_softmax

cake-core/src/backends/vulkan/mod.rs:885–912

Dispatch scaled softmax on GPU: output[row,j] = softmax(input[row,j] * scale). If seq_len > 0, applies causal mask.

(
        &self,
        input: &Tensor,
        rows: usize,
        cols: usize,
        scale: f32,
        seq_len: usize,
    )

    /// Dispatch scaled softmax on GPU: output[row,j] = softmax(input[row,j] * scale).
    /// If seq_len > 0, applies causal mask.
    fn dispatch_softmax(
        &self,
        input: &Tensor,
        rows: usize,
        cols: usize,
        scale: f32,
        seq_len: usize,
    ) -> Result<Tensor> {
        let dtype = input.dtype();
        let shape = input.shape().clone();
        let n = input.elem_count();

        let buf_in = self.get_or_upload(input)?;
        let buf_out = self.alloc_output(n);

        let result = self.dispatch_compute(
            "scaled_softmax",
            &[buf_in.buffer, self.dummy_buf.buffer, buf_out.buffer],
            &buf_out,
            n,
            &[rows as u32, cols as u32, scale.to_bits(), seq_len as u32],
            (rows as u32, 1, 1), // one workgroup per row
        );

        let tensor = Self::from_f32_vec(result, shape.dims(), dtype)?;
        self.cache_activation(tensor.id(), buf_out);
        Ok(tensor)
    }

    // ── GPU matmul ───────────────────────────────────────────────────