| # Integration Guide: Quantum Attention Head |
|
|
| ## 1. Prerequisites |
| - `torch` $\ge 2.0$ |
| - `numpy` $\ge 1.21$ |
| - `cupy` (Optional, for GPU acceleration) |
|
|
| ## 2. Implementation Steps |
|
|
| ### Step 1: QPU Initialization |
| Initialize the `HyperspaceQPU` with the desired qubit count and bond dimension. |
| ```python |
| from core.mini_qpu import HyperspaceQPU |
| qpu = HyperspaceQPU(n_qubits=4, bond_dimension=16) |
| ``` |
|
|
| ### Step 2: Projection Layer |
| Create a linear layer to map the model dimension $d_{\text{model}}$ to the QPU parameter space. |
| ```python |
| self.proj = nn.Linear(d_model, n_qubits * 2) |
| ``` |
| |
| ### Step 3: The Forward Pass |
| Integrate the QPU call within the attention mechanism: |
| 1. Project input to rotation angles. |
| 2. Apply gates to the QPU state. |
| 3. Extract the expectation value as a scalar. |
| 4. Scale the attention output by this scalar. |
| |
| ### Step 4: Block Integration |
| Wrap the `QuantumAttentionHead` within a standard Transformer block, ensuring a `LayerNorm` follows the quantum path to prevent gradient explosion. |
| |
| ## 3. Verification |
| Run the sanity test provided in `core/sovereign_transformer.py`. A successful integration is confirmed when the output shape matches the input shape and the expectation value is non-zero. |
|
|
