tnfr 4.5.2__py3-none-any.whl → 8.5.0__py3-none-any.whl

tnfr/parallel/gpu_engine.py

@@ -0,0 +1,420 @@
+"""GPU acceleration for TNFR computations.
+
+Optional module providing JAX and CuPy integration for GPU-accelerated
+vectorized operations. Requires installation of optional dependencies:
+    pip install tnfr[jax]  # or
+    pip install tnfr[cupy]
+"""
+
+from __future__ import annotations
+
+from typing import Any, Optional
+
+# Check for optional GPU backends
+try:
+    import cupy as cp
+
+    HAS_CUPY = True
+except ImportError:
+    HAS_CUPY = False
+    cp = None  # type: ignore
+
+try:
+    import jax
+    import jax.numpy as jnp
+    from jax import jit
+
+    HAS_JAX = True
+except ImportError:
+    HAS_JAX = False
+    jax = None  # type: ignore
+    jnp = None  # type: ignore
+    jit = None  # type: ignore
+
+
+class TNFRGPUEngine:
+    """GPU acceleration engine for TNFR computations.
+
+    Provides vectorized GPU implementations of ΔNFR and other TNFR operations
+    using JAX or CuPy backends.
+
+    Parameters
+    ----------
+    backend : {"auto", "jax", "cupy", "numpy"}, default="auto"
+        GPU backend to use. "auto" prefers JAX, then CuPy, then NumPy fallback.
+
+    Raises
+    ------
+    ImportError
+        If requested GPU backend is not installed
+
+    Examples
+    --------
+    >>> # Requires JAX or CuPy installation
+    >>> try:
+    ...     from tnfr.parallel import TNFRGPUEngine
+    ...     engine = TNFRGPUEngine(backend="auto")
+    ...     # engine.backend in ["jax", "cupy", "numpy"]
+    ... except ImportError:
+    ...     pass  # Optional dependency not installed
+
+    Notes
+    -----
+    GPU acceleration provides significant speedup for large dense networks
+    but requires compatible hardware and drivers. For sparse networks or
+    small graphs, multiprocessing may be more efficient.
+    """
+
+    def __init__(self, backend: str = "auto"):
+        self.backend = self._select_gpu_backend(backend)
+
+    def _select_gpu_backend(self, backend: str) -> str:
+        """Select available GPU backend."""
+        if backend == "auto":
+            if HAS_JAX:
+                return "jax"
+            elif HAS_CUPY:
+                return "cupy"
+            else:
+                return "numpy"  # Fallback
+
+        if backend == "jax" and not HAS_JAX:
+            raise ImportError("JAX not available. Install with: pip install jax[cuda]")
+        if backend == "cupy" and not HAS_CUPY:
+            raise ImportError("CuPy not available. Install with: pip install cupy")
+
+        return backend
+
+    def compute_delta_nfr_gpu(
+        self,
+        adjacency_matrix: Any,
+        epi_vector: Any,
+        vf_vector: Any,
+        phase_vector: Any,
+    ) -> Any:
+        """Compute ΔNFR using vectorized GPU operations.
+
+        Parameters
+        ----------
+        adjacency_matrix : array-like
+            Network adjacency matrix (N x N)
+        epi_vector : array-like
+            EPI values for all nodes (N,)
+        vf_vector : array-like
+            Structural frequencies νf for all nodes (N,)
+        phase_vector : array-like
+            Phase values θ for all nodes (N,)
+
+        Returns
+        -------
+        array-like
+            ΔNFR values for all nodes (N,)
+
+        Notes
+        -----
+        This is a placeholder for future GPU-accelerated implementations.
+        Actual GPU computation requires careful optimization and testing.
+        Current implementation raises NotImplementedError.
+        """
+        if self.backend == "jax" and HAS_JAX:
+            return self._compute_delta_nfr_jax(
+                adjacency_matrix, epi_vector, vf_vector, phase_vector
+            )
+        elif self.backend == "cupy" and HAS_CUPY:
+            return self._compute_delta_nfr_cupy(
+                adjacency_matrix, epi_vector, vf_vector, phase_vector
+            )
+        else:
+            return self._compute_delta_nfr_numpy(
+                adjacency_matrix, epi_vector, vf_vector, phase_vector
+            )
+
+    def _compute_delta_nfr_jax(
+        self, adj_matrix: Any, epi_vec: Any, vf_vec: Any, phase_vec: Any
+    ) -> Any:
+        """JAX implementation with JIT compilation for GPU acceleration.
+
+        Implements vectorized ΔNFR computation using JAX for automatic
+        GPU acceleration and JIT compilation.
+
+        Parameters
+        ----------
+        adj_matrix : array-like
+            Adjacency matrix (N x N)
+        epi_vec : array-like
+            EPI values (N,)
+        vf_vec : array-like
+            Structural frequencies (N,)
+        phase_vec : array-like
+            Phase values (N,)
+
+        Returns
+        -------
+        jax.numpy.ndarray
+            ΔNFR values for all nodes
+
+        Notes
+        -----
+        Uses the canonical TNFR nodal equation:
+        ∂EPI/∂t = νf · ΔNFR(t)
+
+        ΔNFR is computed from:
+        - Topological gradient (EPI differences with neighbors)
+        - Phase gradient (phase synchronization)
+        - Weighted by structural frequency
+        """
+        if not HAS_JAX:
+            raise ImportError("JAX required for GPU acceleration")
+
+        # Convert inputs to JAX arrays
+        adj = jnp.asarray(adj_matrix)
+        epi = jnp.asarray(epi_vec)
+        vf = jnp.asarray(vf_vec)
+        phase = jnp.asarray(phase_vec)
+
+        # Define JIT-compiled ΔNFR computation
+        @jit
+        def compute_dnfr_vectorized(adj, epi, vf, phase):
+            """Vectorized ΔNFR computation (JIT compiled)."""
+            # Topological gradient: difference in EPI with neighbors
+            # epi_diff[i,j] = epi[j] - epi[i]
+            epi_diff = epi[None, :] - epi[:, None]  # (N, N) matrix
+            topo_gradient = jnp.sum(adj * epi_diff, axis=1)  # (N,) vector
+
+            # Phase gradient: phase difference with neighbors
+            # phase_diff[i,j] = sin(phase[j] - phase[i])
+            phase_diff = jnp.sin(phase[None, :] - phase[:, None])  # (N, N)
+            phase_gradient = jnp.sum(adj * phase_diff, axis=1)  # (N,)
+
+            # Normalize by degree (number of neighbors)
+            degree = jnp.sum(adj, axis=1)
+            # Avoid division by zero
+            degree_safe = jnp.where(degree > 0, degree, 1.0)
+
+            topo_gradient = topo_gradient / degree_safe
+            phase_gradient = phase_gradient / degree_safe
+
+            # Combine gradients with TNFR weights
+            # Emphasize topological structure (0.7) over phase (0.3)
+            combined_gradient = 0.7 * topo_gradient + 0.3 * phase_gradient
+
+            # Apply structural frequency modulation (canonical equation)
+            delta_nfr = vf * combined_gradient
+
+            return delta_nfr
+
+        # Execute JIT-compiled computation (GPU accelerated if available)
+        result = compute_dnfr_vectorized(adj, epi, vf, phase)
+
+        return result
+
+    def _compute_delta_nfr_cupy(
+        self, adj_matrix: Any, epi_vec: Any, vf_vec: Any, phase_vec: Any
+    ) -> Any:
+        """CuPy implementation for CUDA GPUs.
+
+        Implements vectorized ΔNFR computation using CuPy for CUDA GPU
+        acceleration with NumPy-compatible interface.
+
+        Parameters
+        ----------
+        adj_matrix : array-like
+            Adjacency matrix (N x N)
+        epi_vec : array-like
+            EPI values (N,)
+        vf_vec : array-like
+            Structural frequencies (N,)
+        phase_vec : array-like
+            Phase values (N,)
+
+        Returns
+        -------
+        cupy.ndarray
+            ΔNFR values for all nodes (on GPU)
+        """
+        if not HAS_CUPY:
+            raise ImportError("CuPy required for CUDA GPU acceleration")
+
+        # Transfer to GPU
+        adj = cp.asarray(adj_matrix)
+        epi = cp.asarray(epi_vec)
+        vf = cp.asarray(vf_vec)
+        phase = cp.asarray(phase_vec)
+
+        # Topological gradient (vectorized on GPU)
+        epi_diff = epi[None, :] - epi[:, None]
+        topo_gradient = cp.sum(adj * epi_diff, axis=1)
+
+        # Phase gradient (vectorized on GPU)
+        phase_diff = cp.sin(phase[None, :] - phase[:, None])
+        phase_gradient = cp.sum(adj * phase_diff, axis=1)
+
+        # Normalize by degree
+        degree = cp.sum(adj, axis=1)
+        degree_safe = cp.where(degree > 0, degree, 1.0)
+
+        topo_gradient = topo_gradient / degree_safe
+        phase_gradient = phase_gradient / degree_safe
+
+        # Combine with TNFR weights
+        combined_gradient = 0.7 * topo_gradient + 0.3 * phase_gradient
+
+        # Apply structural frequency
+        delta_nfr = vf * combined_gradient
+
+        return delta_nfr
+
+    def _compute_delta_nfr_numpy(
+        self, adj_matrix: Any, epi_vec: Any, vf_vec: Any, phase_vec: Any
+    ) -> Any:
+        """NumPy fallback implementation (CPU-only).
+
+        Provides CPU-based vectorized computation when GPU is unavailable.
+
+        Parameters
+        ----------
+        adj_matrix : array-like
+            Adjacency matrix (N x N)
+        epi_vec : array-like
+            EPI values (N,)
+        vf_vec : array-like
+            Structural frequencies (N,)
+        phase_vec : array-like
+            Phase values (N,)
+
+        Returns
+        -------
+        numpy.ndarray
+            ΔNFR values for all nodes
+        """
+        try:
+            import numpy as np
+        except ImportError:
+            raise ImportError("NumPy required for CPU computation")
+
+        # Convert to numpy arrays
+        adj = np.asarray(adj_matrix)
+        epi = np.asarray(epi_vec)
+        vf = np.asarray(vf_vec)
+        phase = np.asarray(phase_vec)
+
+        # Topological gradient
+        epi_diff = epi[None, :] - epi[:, None]
+        topo_gradient = np.sum(adj * epi_diff, axis=1)
+
+        # Phase gradient
+        phase_diff = np.sin(phase[None, :] - phase[:, None])
+        phase_gradient = np.sum(adj * phase_diff, axis=1)
+
+        # Normalize by degree
+        degree = np.sum(adj, axis=1)
+        degree_safe = np.where(degree > 0, degree, 1.0)
+
+        topo_gradient = topo_gradient / degree_safe
+        phase_gradient = phase_gradient / degree_safe
+
+        # Combine with TNFR weights
+        combined_gradient = 0.7 * topo_gradient + 0.3 * phase_gradient
+
+        # Apply structural frequency
+        delta_nfr = vf * combined_gradient
+
+        return delta_nfr
+
+    def compute_delta_nfr_from_graph(self, graph: Any) -> Dict[Any, float]:
+        """Compute ΔNFR directly from a TNFR graph using GPU acceleration.
+
+        Convenience method that extracts matrices from graph and computes
+        ΔNFR using GPU backend.
+
+        Parameters
+        ----------
+        graph : TNFRGraph
+            Network graph with TNFR attributes
+
+        Returns
+        -------
+        Dict[Any, float]
+            Mapping from node IDs to ΔNFR values
+
+        Examples
+        --------
+        >>> import networkx as nx
+        >>> from tnfr.parallel import TNFRGPUEngine
+        >>> G = nx.Graph([(0, 1), (1, 2)])
+        >>> for node in G.nodes():
+        ...     G.nodes[node]['epi'] = 0.5
+        ...     G.nodes[node]['nu_f'] = 1.0
+        ...     G.nodes[node]['phase'] = 0.0
+        >>> engine = TNFRGPUEngine(backend="numpy")  # Use numpy for testing
+        >>> result = engine.compute_delta_nfr_from_graph(G)
+        >>> len(result) == 3
+        True
+        """
+        import networkx as nx
+
+        try:
+            import numpy as np
+        except ImportError:
+            raise ImportError("NumPy required for graph processing")
+
+        # Extract node list (maintain order)
+        nodes = list(graph.nodes())
+        node_to_idx = {node: idx for idx, node in enumerate(nodes)}
+
+        # Build adjacency matrix
+        n = len(nodes)
+        adj_matrix = np.zeros((n, n))
+        for i, j in graph.edges():
+            idx_i = node_to_idx[i]
+            idx_j = node_to_idx[j]
+            adj_matrix[idx_i, idx_j] = 1.0
+            adj_matrix[idx_j, idx_i] = 1.0  # Undirected
+
+        # Extract node attributes
+        def get_attr(node, attr_names, default):
+            """Get attribute with fallbacks."""
+            for name in (
+                attr_names if isinstance(attr_names, (list, tuple)) else [attr_names]
+            ):
+                if name in graph.nodes[node]:
+                    return float(graph.nodes[node][name])
+            return default
+
+        epi_vec = np.array([get_attr(node, ["epi", "EPI"], 0.5) for node in nodes])
+        vf_vec = np.array([get_attr(node, ["nu_f", "vf", "νf"], 1.0) for node in nodes])
+        phase_vec = np.array(
+            [get_attr(node, ["phase", "theta"], 0.0) for node in nodes]
+        )
+
+        # Compute ΔNFR using GPU
+        delta_nfr_array = self.compute_delta_nfr_gpu(
+            adj_matrix, epi_vec, vf_vec, phase_vec
+        )
+
+        # Convert back to dictionary
+        if self.backend == "cupy" and HAS_CUPY:
+            delta_nfr_array = cp.asnumpy(delta_nfr_array)  # Transfer from GPU
+        elif self.backend == "jax" and HAS_JAX:
+            delta_nfr_array = np.array(delta_nfr_array)  # Convert from JAX
+
+        result = {node: float(delta_nfr_array[idx]) for idx, node in enumerate(nodes)}
+
+        return result
+
+    @property
+    def is_gpu_available(self) -> bool:
+        """Check if GPU acceleration is actually available."""
+        if self.backend == "jax" and HAS_JAX:
+            try:
+                # Check if JAX has GPU backend
+                return len(jax.devices("gpu")) > 0
+            except Exception:
+                return False
+        elif self.backend == "cupy" and HAS_CUPY:
+            try:
+                # Check if CuPy can access GPU
+                return cp.cuda.runtime.getDeviceCount() > 0
+            except Exception:
+                return False
+        return False

tnfr/parallel/monitoring.py

@@ -0,0 +1,248 @@
+"""Performance monitoring for parallel TNFR computations.
+
+Tracks execution metrics to enable optimization and auto-scaling decisions.
+"""
+
+from __future__ import annotations
+
+import time
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional
+
+try:
+    import psutil
+
+    HAS_PSUTIL = True
+except ImportError:
+    HAS_PSUTIL = False
+
+
+@dataclass
+class PerformanceMetrics:
+    """Performance metrics for parallel TNFR execution.
+
+    Attributes
+    ----------
+    start_time : float
+        Unix timestamp when execution started
+    end_time : float
+        Unix timestamp when execution completed
+    duration_seconds : float
+        Total execution time in seconds
+    peak_memory_mb : float
+        Peak memory usage in megabytes
+    avg_cpu_percent : float
+        Average CPU utilization percentage
+    workers_used : int
+        Number of parallel workers employed
+    nodes_processed : int
+        Total number of nodes processed
+    operations_per_second : float
+        Throughput metric (nodes/second)
+    coherence_improvement : float
+        Change in global coherence C(t)
+    parallelization_efficiency : float
+        Actual speedup / theoretical speedup ratio
+    memory_efficiency : float
+        Useful work / total memory ratio
+    """
+
+    start_time: float
+    end_time: float
+    duration_seconds: float
+    peak_memory_mb: float
+    avg_cpu_percent: float
+    workers_used: int
+    nodes_processed: int
+    operations_per_second: float
+    coherence_improvement: float
+    parallelization_efficiency: float
+    memory_efficiency: float
+
+
+class ParallelExecutionMonitor:
+    """Real-time monitoring for parallel TNFR execution.
+
+    Tracks resource usage, throughput, and efficiency metrics during parallel
+    computation to enable dynamic optimization and post-execution analysis.
+
+    Examples
+    --------
+    >>> from tnfr.parallel import ParallelExecutionMonitor
+    >>> monitor = ParallelExecutionMonitor()
+    >>> monitor.start_monitoring(expected_nodes=100, workers=2)
+    >>> # ... perform computation ...
+    >>> metrics = monitor.stop_monitoring(
+    ...     final_coherence=0.85,
+    ...     initial_coherence=0.75
+    ... )
+    >>> metrics.nodes_processed
+    100
+    >>> metrics.workers_used
+    2
+    """
+
+    def __init__(self):
+        self._metrics_history: List[PerformanceMetrics] = []
+        self._current_metrics: Optional[Dict[str, Any]] = None
+        self._process = None
+        if HAS_PSUTIL:
+            try:
+                import psutil
+
+                self._process = psutil.Process()
+            except Exception:
+                self._process = None
+
+    def start_monitoring(self, expected_nodes: int, workers: int) -> None:
+        """Start monitoring execution.
+
+        Parameters
+        ----------
+        expected_nodes : int
+            Expected number of nodes to process
+        workers : int
+            Number of parallel workers
+        """
+        self._current_metrics = {
+            "start_time": time.time(),
+            "expected_nodes": expected_nodes,
+            "workers": workers,
+            "memory_samples": [],
+            "cpu_samples": [],
+        }
+
+        # Take initial resource snapshot
+        if self._process:
+            try:
+                mem_info = self._process.memory_info()
+                self._current_metrics["memory_samples"].append(
+                    mem_info.rss / 1024 / 1024
+                )
+                self._current_metrics["cpu_samples"].append(self._process.cpu_percent())
+            except Exception:
+                pass
+
+    def stop_monitoring(
+        self, final_coherence: float, initial_coherence: float
+    ) -> PerformanceMetrics:
+        """Stop monitoring and compute final metrics.
+
+        Parameters
+        ----------
+        final_coherence : float
+            Final network coherence C(t)
+        initial_coherence : float
+            Initial network coherence C(t)
+
+        Returns
+        -------
+        PerformanceMetrics
+            Complete performance metrics for the execution
+        """
+        if self._current_metrics is None:
+            raise RuntimeError("Monitoring not started")
+
+        end_time = time.time()
+        duration = end_time - self._current_metrics["start_time"]
+
+        # Take final resource snapshot
+        if self._process:
+            try:
+                mem_info = self._process.memory_info()
+                self._current_metrics["memory_samples"].append(
+                    mem_info.rss / 1024 / 1024
+                )
+                self._current_metrics["cpu_samples"].append(self._process.cpu_percent())
+            except Exception:
+                pass
+
+        # Calculate aggregated metrics
+        memory_samples = self._current_metrics.get("memory_samples", [])
+        cpu_samples = self._current_metrics.get("cpu_samples", [])
+
+        peak_memory = max(memory_samples) if memory_samples else 0.0
+        avg_cpu = sum(cpu_samples) / len(cpu_samples) if cpu_samples else 0.0
+
+        nodes = self._current_metrics["expected_nodes"]
+        workers = self._current_metrics["workers"]
+
+        # Calculate parallelization efficiency
+        # NOTE: This is a heuristic approximation. True efficiency requires:
+        # - Baseline sequential measurement
+        # - Accounting for Amdahl's law (sequential portions)
+        # - Consideration of communication overhead
+        # Current approach: estimate from CPU utilization as proxy
+        theoretical_speedup = workers
+        # Estimate actual speedup from CPU utilization
+        # If we're using N workers, we expect ~N * 100% CPU in ideal case
+        expected_cpu = workers * 100.0
+        actual_speedup = (avg_cpu / 100.0) if expected_cpu > 0 else 1.0
+        parallelization_eff = (
+            min(1.0, actual_speedup / theoretical_speedup)
+            if theoretical_speedup > 0
+            else 0.0
+        )
+
+        # Memory efficiency: nodes per MB
+        memory_eff = nodes / peak_memory if peak_memory > 0 else 0.0
+
+        metrics = PerformanceMetrics(
+            start_time=self._current_metrics["start_time"],
+            end_time=end_time,
+            duration_seconds=duration,
+            peak_memory_mb=peak_memory,
+            avg_cpu_percent=avg_cpu,
+            workers_used=workers,
+            nodes_processed=nodes,
+            operations_per_second=nodes / duration if duration > 0 else 0.0,
+            coherence_improvement=final_coherence - initial_coherence,
+            parallelization_efficiency=parallelization_eff,
+            memory_efficiency=memory_eff,
+        )
+
+        self._metrics_history.append(metrics)
+        self._current_metrics = None
+
+        return metrics
+
+    def get_optimization_suggestions(self) -> List[str]:
+        """Generate optimization suggestions based on execution history.
+
+        Returns
+        -------
+        List[str]
+            List of actionable suggestions for improving performance
+        """
+        if not self._metrics_history:
+            return ["No execution history available"]
+
+        latest = self._metrics_history[-1]
+        suggestions = []
+
+        if latest.parallelization_efficiency < 0.5:
+            suggestions.append(
+                "⚡ Low parallelization efficiency - consider reducing "
+                "worker count or increasing chunk size"
+            )
+
+        if latest.memory_efficiency < 0.1:
+            suggestions.append(
+                "💾 High memory usage - consider distributed execution "
+                "or memory optimization"
+            )
+
+        if latest.operations_per_second < 100:
+            suggestions.append(
+                "📈 Low throughput - consider GPU backend or algorithm " "optimization"
+            )
+
+        if not suggestions:
+            suggestions.append("✨ Performance looks optimal!")
+
+        return suggestions
+
+    @property
+    def history(self) -> List[PerformanceMetrics]:
+        """Get execution history."""
+        return self._metrics_history.copy()