vectorphile 0.2.0__tar.gz → 0.3.0__tar.gz

{vectorphile-0.2.0/vectorphile.egg-info → vectorphile-0.3.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: vectorphile
-Version: 0.2.0
+Version: 0.3.0
 Summary: Unified ML/DL Python Library — replaces PyTorch + scikit-learn + NumPy + Pandas + SciPy + XGBoost + LightGBM + CatBoost + JAX + Matplotlib + Seaborn
 Home-page: https://github.com/vectorphile/vectorphile
 Author: vectorphile
@@ -19,12 +19,37 @@ Requires-Dist: h5py>=3.8.0; extra == "io"
 Requires-Dist: netCDF4>=1.6.0; extra == "io"
 Provides-Extra: viz
 Requires-Dist: matplotlib>=3.7.0; extra == "viz"
+Provides-Extra: cuda
+Requires-Dist: cupy-cuda12x>=12.0.0; extra == "cuda"
+Provides-Extra: numba
+Requires-Dist: numba>=0.57.0; extra == "numba"
+Provides-Extra: triton
+Requires-Dist: triton>=2.1.0; extra == "triton"
+Provides-Extra: rl
+Requires-Dist: gymnasium>=0.29.0; extra == "rl"
+Provides-Extra: tokenizers
+Requires-Dist: tokenizers>=0.15.0; extra == "tokenizers"
+Provides-Extra: onnx
+Requires-Dist: onnx>=1.14.0; extra == "onnx"
+Requires-Dist: onnxruntime>=1.16.0; extra == "onnx"
+Provides-Extra: coreml
+Requires-Dist: coremltools>=7.0; extra == "coreml"
+Provides-Extra: pretrained
+Requires-Dist: huggingface-hub>=0.20.0; extra == "pretrained"
+Requires-Dist: safetensors>=0.4.0; extra == "pretrained"
 Provides-Extra: full
 Requires-Dist: scipy>=1.10.0; extra == "full"
 Requires-Dist: pyarrow>=12.0.0; extra == "full"
 Requires-Dist: h5py>=3.8.0; extra == "full"
 Requires-Dist: netCDF4>=1.6.0; extra == "full"
 Requires-Dist: matplotlib>=3.7.0; extra == "full"
+Requires-Dist: numba>=0.57.0; extra == "full"
+Requires-Dist: triton>=2.1.0; extra == "full"
+Requires-Dist: gymnasium>=0.29.0; extra == "full"
+Requires-Dist: onnx>=1.14.0; extra == "full"
+Requires-Dist: onnxruntime>=1.16.0; extra == "full"
+Requires-Dist: huggingface-hub>=0.20.0; extra == "full"
+Requires-Dist: safetensors>=0.4.0; extra == "full"
 Provides-Extra: dev
 Requires-Dist: pytest>=7.0.0; extra == "dev"
 Requires-Dist: pytest-cov>=4.0.0; extra == "dev"
@@ -101,6 +126,26 @@ This is not a wrapper. Every module is a ground-up implementation, designed with
 | `vp.quantization` | quantize (INT8/INT4/FP16), QuantizedModel, MinMaxCalibrator, magnitude_prune |
 | `vp.amp` | autocast, GradScaler (Phase 2: real CUDA fp16/bf16) |
 | `vp.utils.data` | `Dataset`, `TensorDataset`, `DataLoader` with batching and shuffling |
+| `vp.backend` | 5-tier dispatch: Triton JIT → PyTorch CUDA → Numba CUDA → im2col+BLAS → NumPy |
+| `vp.tokenizers` | BPE, WordPiece, SentencePiece, CharacterTokenizer — zero external deps |
+| `vp.distributed` | DataParallel, DDP, GradientAccumulator, MixedPrecisionTrainer |
+| `vp.export` | ONNX, TorchScript, CoreML, SavedModel, GGUF (Q4_0/Q8_0), benchmark |
+| `vp.rl` | DQN/DoubleDQN/DuelingDQN, REINFORCE, A2C, PPO, SAC, TD3, PER, environments |
+| `vp.pretrained` | HuggingFace Hub download, safetensors parser, weight mappers for GPT-2/BERT/ViT/LLaMA |
+| `vp.metrics` | 25+ metrics: accuracy, BLEU, ROUGE-L, WER, NDCG, ECE, silhouette, ARI, NMI |
+| `vp.callbacks` | EarlyStopping, ModelCheckpoint, ReduceLROnPlateau, TensorBoardLogger, WandBLogger, LRFinder |
+| `vp.trainer` | PyTorch Lightning-style `Trainer` + `TrainableModel` base class |
+| `vp.text` | beam_search, nucleus_sampling, contrastive_search, speculative_decoding, chat templates, `TextGenerationPipeline` |
+| `vp.vision` | Compose, Resize, RandomCrop, ColorJitter, AugMix, ToTensor, Normalize, make_grid, extract_features |
+| `vp.data.datasets` | MNIST, CIFAR-10, TinyShakespeare, IMDb + DataLoader, samplers, random_split |
+| `vp.nn.lr_schedulers` | StepLR, CosineAnnealingLR, OneCycleLR, WarmupCosineDecay, PolynomialLR, get_scheduler |
+| `vp.nn.einops` | rearrange, reduce, repeat, pack, unpack, EinMix |
+| `vp.nn.hooks` | register_forward_hook, register_backward_hook, FeatureExtractor, GradientRecorder |
+| `vp.autograd` | Reverse-mode + second-order autodiff, `jacobian`, `hessian`, `vmap`, `checkpoint` |
+| `vp.nn.flash_attention` | Flash Attention 2 (tiled, O(N) memory), variable-length sequences |
+| `vp.nn.qat` | Quantization-aware training (straight-through estimator), prepare_qat / convert_qat |
+| `vp.optim` (advanced) | Lion, Sophia, Adan, CAME, Prodigy, ScheduleFreeAdamW |
+| `vp.nn.architectures` | CNN zoo (VGG/ResNet/DenseNet/EfficientNet), ViT/DeiT, Mamba/S4, VAE/DDPM/DCGAN, U-Net, GCN/GAT/GIN |
 
 ---
 
@@ -451,26 +496,40 @@ vectorphile/
 
 ## Roadmap
 
-### Phase 1 — Pure Python (current)
+### v0.1–0.2 — Core (complete)
 - Full tensor API with reverse-mode autodiff
-- Complete neural network layer set
-- All classical ML algorithms
-- DataFrame with full GroupBy/merge/pivot
+- Complete neural network layer set (Linear, Conv, RNN, Transformer, Norm)
+- All classical ML algorithms (trees, forests, XGBoost, clustering, PCA)
+- DataFrame, GroupBy, merge/pivot
 - Linear algebra, statistics, FFT, signal, sparse, optimization
 
-### Phase 2 — C + CUDA Kernels
-- Custom C extensions for core tensor ops via CPython API
-- Real CUDA support: cuBLAS for GEMM, cuDNN for convolutions
-- fp16 / bf16 automatic mixed precision
-- cuFFT backend for spectral ops
-- Triton kernels for custom GPU operations
-
-### Phase 3 — Distributed + Production
-- Multi-GPU data parallelism
-- Gradient checkpointing for memory-efficient backprop
-- ONNX export
-- Quantization (int8, int4)
-- JIT compilation via traced execution graphs
+### v0.3.0 — Production completeness (complete)
+- 5-tier CUDA backend dispatcher (Triton → PyTorch → Numba → im2col → NumPy)
+- BPE / WordPiece / SentencePiece tokenizers from scratch
+- Distributed training: DataParallel, DDP, MixedPrecision
+- Model export: ONNX, TorchScript, CoreML
+- Reinforcement learning: DQN, PPO, SAC, TD3
+- Pretrained weights hub (HuggingFace download, safetensors parser)
+- Architecture zoo: CNN/ResNet/EfficientNet, ViT/DeiT, Mamba/S4, VAE/DDPM, U-Net, GCN/GAT/GIN
+- Flash Attention, QAT, advanced optimizers (Lion, Sophia, Adan, Prodigy)
+
+### v0.4.0 — Ecosystem parity (complete)
+- GGUF binary format reader/writer (Q4_0, Q8_0 quantization)
+- 25+ evaluation metrics (BLEU, ROUGE, WER, ECE, NDCG, silhouette)
+- 10 training callbacks (EarlyStopping, TensorBoardLogger, WandBLogger, LRFinder)
+- Keras/Lightning-style `Trainer` + `TrainableModel` base class
+- 11 LR schedulers (`get_scheduler` factory, WarmupCosineDecay, OneCycleLR)
+- einops-compatible `rearrange`/`reduce`/`repeat`/`pack`/`unpack`/`EinMix`
+- Forward/backward hooks, `FeatureExtractor`, `GradientRecorder`
+- Text generation: beam search, nucleus, contrastive, speculative decoding, chat templates
+- Vision transforms: Compose/Resize/ColorJitter/AugMix/Normalize + ImageNet defaults
+- Datasets: MNIST, CIFAR-10, TinyShakespeare, IMDb + DataLoader/samplers
+
+### v0.5.0 — Planned
+- Real CUDA C extensions (cuBLAS GEMM, cuDNN conv)
+- fp16 / bf16 automatic mixed precision on GPU
+- Triton fused kernels in production paths
+- Multi-node distributed via NCCL
 
 ---
 

vectorphile-0.2.0/PKG-INFO → vectorphile-0.3.0/README.md

@@ -1,39 +1,3 @@
-Metadata-Version: 2.4
-Name: vectorphile
-Version: 0.2.0
-Summary: Unified ML/DL Python Library — replaces PyTorch + scikit-learn + NumPy + Pandas + SciPy + XGBoost + LightGBM + CatBoost + JAX + Matplotlib + Seaborn
-Home-page: https://github.com/vectorphile/vectorphile
-Author: vectorphile
-Author-email: vectorphile <saadsalmanakram1@gmail.com>
-License: Apache-2.0
-Project-URL: Homepage, https://github.com/vectorphile/vectorphile
-Project-URL: Bug Tracker, https://github.com/vectorphile/vectorphile/issues
-Requires-Python: >=3.9
-Description-Content-Type: text/markdown
-License-File: LICENSE
-Requires-Dist: numpy>=1.24.0
-Requires-Dist: scipy>=1.10.0
-Provides-Extra: io
-Requires-Dist: pyarrow>=12.0.0; extra == "io"
-Requires-Dist: h5py>=3.8.0; extra == "io"
-Requires-Dist: netCDF4>=1.6.0; extra == "io"
-Provides-Extra: viz
-Requires-Dist: matplotlib>=3.7.0; extra == "viz"
-Provides-Extra: full
-Requires-Dist: scipy>=1.10.0; extra == "full"
-Requires-Dist: pyarrow>=12.0.0; extra == "full"
-Requires-Dist: h5py>=3.8.0; extra == "full"
-Requires-Dist: netCDF4>=1.6.0; extra == "full"
-Requires-Dist: matplotlib>=3.7.0; extra == "full"
-Provides-Extra: dev
-Requires-Dist: pytest>=7.0.0; extra == "dev"
-Requires-Dist: pytest-cov>=4.0.0; extra == "dev"
-Requires-Dist: ruff>=0.1.0; extra == "dev"
-Dynamic: author
-Dynamic: home-page
-Dynamic: license-file
-Dynamic: requires-python
-
 # vectorphile
 
 > **A unified, research-first computational intelligence library for Python.**
@@ -101,6 +65,26 @@ This is not a wrapper. Every module is a ground-up implementation, designed with
 | `vp.quantization` | quantize (INT8/INT4/FP16), QuantizedModel, MinMaxCalibrator, magnitude_prune |
 | `vp.amp` | autocast, GradScaler (Phase 2: real CUDA fp16/bf16) |
 | `vp.utils.data` | `Dataset`, `TensorDataset`, `DataLoader` with batching and shuffling |
+| `vp.backend` | 5-tier dispatch: Triton JIT → PyTorch CUDA → Numba CUDA → im2col+BLAS → NumPy |
+| `vp.tokenizers` | BPE, WordPiece, SentencePiece, CharacterTokenizer — zero external deps |
+| `vp.distributed` | DataParallel, DDP, GradientAccumulator, MixedPrecisionTrainer |
+| `vp.export` | ONNX, TorchScript, CoreML, SavedModel, GGUF (Q4_0/Q8_0), benchmark |
+| `vp.rl` | DQN/DoubleDQN/DuelingDQN, REINFORCE, A2C, PPO, SAC, TD3, PER, environments |
+| `vp.pretrained` | HuggingFace Hub download, safetensors parser, weight mappers for GPT-2/BERT/ViT/LLaMA |
+| `vp.metrics` | 25+ metrics: accuracy, BLEU, ROUGE-L, WER, NDCG, ECE, silhouette, ARI, NMI |
+| `vp.callbacks` | EarlyStopping, ModelCheckpoint, ReduceLROnPlateau, TensorBoardLogger, WandBLogger, LRFinder |
+| `vp.trainer` | PyTorch Lightning-style `Trainer` + `TrainableModel` base class |
+| `vp.text` | beam_search, nucleus_sampling, contrastive_search, speculative_decoding, chat templates, `TextGenerationPipeline` |
+| `vp.vision` | Compose, Resize, RandomCrop, ColorJitter, AugMix, ToTensor, Normalize, make_grid, extract_features |
+| `vp.data.datasets` | MNIST, CIFAR-10, TinyShakespeare, IMDb + DataLoader, samplers, random_split |
+| `vp.nn.lr_schedulers` | StepLR, CosineAnnealingLR, OneCycleLR, WarmupCosineDecay, PolynomialLR, get_scheduler |
+| `vp.nn.einops` | rearrange, reduce, repeat, pack, unpack, EinMix |
+| `vp.nn.hooks` | register_forward_hook, register_backward_hook, FeatureExtractor, GradientRecorder |
+| `vp.autograd` | Reverse-mode + second-order autodiff, `jacobian`, `hessian`, `vmap`, `checkpoint` |
+| `vp.nn.flash_attention` | Flash Attention 2 (tiled, O(N) memory), variable-length sequences |
+| `vp.nn.qat` | Quantization-aware training (straight-through estimator), prepare_qat / convert_qat |
+| `vp.optim` (advanced) | Lion, Sophia, Adan, CAME, Prodigy, ScheduleFreeAdamW |
+| `vp.nn.architectures` | CNN zoo (VGG/ResNet/DenseNet/EfficientNet), ViT/DeiT, Mamba/S4, VAE/DDPM/DCGAN, U-Net, GCN/GAT/GIN |
 
 ---
 
@@ -451,26 +435,40 @@ vectorphile/
 
 ## Roadmap
 
-### Phase 1 — Pure Python (current)
+### v0.1–0.2 — Core (complete)
 - Full tensor API with reverse-mode autodiff
-- Complete neural network layer set
-- All classical ML algorithms
-- DataFrame with full GroupBy/merge/pivot
+- Complete neural network layer set (Linear, Conv, RNN, Transformer, Norm)
+- All classical ML algorithms (trees, forests, XGBoost, clustering, PCA)
+- DataFrame, GroupBy, merge/pivot
 - Linear algebra, statistics, FFT, signal, sparse, optimization
 
-### Phase 2 — C + CUDA Kernels
-- Custom C extensions for core tensor ops via CPython API
-- Real CUDA support: cuBLAS for GEMM, cuDNN for convolutions
-- fp16 / bf16 automatic mixed precision
-- cuFFT backend for spectral ops
-- Triton kernels for custom GPU operations
-
-### Phase 3 — Distributed + Production
-- Multi-GPU data parallelism
-- Gradient checkpointing for memory-efficient backprop
-- ONNX export
-- Quantization (int8, int4)
-- JIT compilation via traced execution graphs
+### v0.3.0 — Production completeness (complete)
+- 5-tier CUDA backend dispatcher (Triton → PyTorch → Numba → im2col → NumPy)
+- BPE / WordPiece / SentencePiece tokenizers from scratch
+- Distributed training: DataParallel, DDP, MixedPrecision
+- Model export: ONNX, TorchScript, CoreML
+- Reinforcement learning: DQN, PPO, SAC, TD3
+- Pretrained weights hub (HuggingFace download, safetensors parser)
+- Architecture zoo: CNN/ResNet/EfficientNet, ViT/DeiT, Mamba/S4, VAE/DDPM, U-Net, GCN/GAT/GIN
+- Flash Attention, QAT, advanced optimizers (Lion, Sophia, Adan, Prodigy)
+
+### v0.4.0 — Ecosystem parity (complete)
+- GGUF binary format reader/writer (Q4_0, Q8_0 quantization)
+- 25+ evaluation metrics (BLEU, ROUGE, WER, ECE, NDCG, silhouette)
+- 10 training callbacks (EarlyStopping, TensorBoardLogger, WandBLogger, LRFinder)
+- Keras/Lightning-style `Trainer` + `TrainableModel` base class
+- 11 LR schedulers (`get_scheduler` factory, WarmupCosineDecay, OneCycleLR)
+- einops-compatible `rearrange`/`reduce`/`repeat`/`pack`/`unpack`/`EinMix`
+- Forward/backward hooks, `FeatureExtractor`, `GradientRecorder`
+- Text generation: beam search, nucleus, contrastive, speculative decoding, chat templates
+- Vision transforms: Compose/Resize/ColorJitter/AugMix/Normalize + ImageNet defaults
+- Datasets: MNIST, CIFAR-10, TinyShakespeare, IMDb + DataLoader/samplers
+
+### v0.5.0 — Planned
+- Real CUDA C extensions (cuBLAS GEMM, cuDNN conv)
+- fp16 / bf16 automatic mixed precision on GPU
+- Triton fused kernels in production paths
+- Multi-node distributed via NCCL
 
 ---
 

{vectorphile-0.2.0 → vectorphile-0.3.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "vectorphile"
-version = "0.2.0"
+version = "0.3.0"
 description = "Unified ML/DL Python Library — replaces PyTorch + scikit-learn + NumPy + Pandas + SciPy + XGBoost + LightGBM + CatBoost + JAX + Matplotlib + Seaborn"
 readme = "README.md"
 license = { text = "Apache-2.0" }
@@ -18,12 +18,35 @@ dependencies = [
 [project.optional-dependencies]
 io = ["pyarrow>=12.0.0", "h5py>=3.8.0", "netCDF4>=1.6.0"]
 viz = ["matplotlib>=3.7.0"]
+# GPU acceleration via CuPy (pick the right CUDA variant for your driver)
+cuda = ["cupy-cuda12x>=12.0.0"]
+# Numba CUDA kernels — custom GPU kernels, no PyTorch required
+numba = ["numba>=0.57.0"]
+# OpenAI Triton fused kernels — maximum GPU throughput (requires torch + CUDA)
+triton = ["triton>=2.1.0"]
+# Reinforcement learning extras (compatible environment wrappers)
+rl = ["gymnasium>=0.29.0"]
+# Tokenizer performance (optional C extension for BPE)
+tokenizers = ["tokenizers>=0.15.0"]
+# ONNX export and runtime
+onnx = ["onnx>=1.14.0", "onnxruntime>=1.16.0"]
+# CoreML export (macOS)
+coreml = ["coremltools>=7.0"]
+# Pretrained model hub (huggingface-hub is optional — urllib used by default)
+pretrained = ["huggingface-hub>=0.20.0", "safetensors>=0.4.0"]
 full = [
     "scipy>=1.10.0",
     "pyarrow>=12.0.0",
     "h5py>=3.8.0",
     "netCDF4>=1.6.0",
     "matplotlib>=3.7.0",
+    "numba>=0.57.0",
+    "triton>=2.1.0",
+    "gymnasium>=0.29.0",
+    "onnx>=1.14.0",
+    "onnxruntime>=1.16.0",
+    "huggingface-hub>=0.20.0",
+    "safetensors>=0.4.0",
 ]
 dev = ["pytest>=7.0.0", "pytest-cov>=4.0.0", "ruff>=0.1.0"]
 

{vectorphile-0.2.0 → vectorphile-0.3.0}/vectorphile/__init__.py

@@ -28,7 +28,7 @@ Quick start:
     df.groupby('category').mean()
 """
 
-__version__ = '0.2.0'
+__version__ = '0.3.0'
 __author__ = 'vectorphile'
 __license__ = 'Apache-2.0'
 
@@ -107,6 +107,22 @@ from vectorphile import quantization
 from vectorphile import utils
 from vectorphile.utils.data import Dataset, TensorDataset, DataLoader
 
+# ── New modules (v0.3.0 — production completeness) ───────────────────────────
+from vectorphile import backend
+from vectorphile import tokenizers
+from vectorphile import distributed
+from vectorphile import export
+from vectorphile import rl
+from vectorphile import pretrained
+
+# ── New modules (v0.4.0 — full ecosystem parity) ─────────────────────────────
+from vectorphile import metrics
+from vectorphile import callbacks
+from vectorphile import trainer
+from vectorphile import text
+from vectorphile import vision
+from vectorphile.data import datasets
+
 # ── Convenience re-exports ───────────────────────────────────────────────────
 from vectorphile.data import DataFrame, Series, read_csv, read_json
 from vectorphile.ml import (
@@ -157,6 +173,8 @@ __all__ = [
     'data', 'ml', 'utils',
     'integrate', 'interpolate', 'spatial', 'ndimage', 'transforms',
     'polynomial', 'io', 'ma', 'viz', 'hub', 'quantization',
+    'backend', 'tokenizers', 'distributed', 'export', 'rl', 'pretrained',
+    'metrics', 'callbacks', 'trainer', 'text', 'vision', 'datasets',
     'Dataset', 'TensorDataset', 'DataLoader',
     'DataFrame', 'Series', 'read_csv', 'read_json',
     'LinearRegression', 'Ridge', 'Lasso', 'ElasticNet', 'LogisticRegression',

{vectorphile-0.2.0 → vectorphile-0.3.0}/vectorphile/autograd/__init__.py

@@ -4,10 +4,16 @@ from vectorphile.autograd.engine import (
     ExpBackward, LogBackward, SqrtBackward, SumBackward, MeanBackward,
     ReLUBackward, SigmoidBackward, TanhBackward,
 )
+from vectorphile.autograd.functional import (
+    grad, jacobian, hessian, vmap, checkpoint, enable_grad,
+    zeros, ones, randn, cat, stack,
+)
 
 __all__ = [
-    'backward', 'no_grad',
+    'backward', 'no_grad', 'enable_grad',
     'AddBackward', 'SubBackward', 'MulBackward', 'DivBackward', 'MatMulBackward',
     'ExpBackward', 'LogBackward', 'SqrtBackward', 'SumBackward', 'MeanBackward',
     'ReLUBackward', 'SigmoidBackward', 'TanhBackward',
+    'grad', 'jacobian', 'hessian', 'vmap', 'checkpoint',
+    'zeros', 'ones', 'randn', 'cat', 'stack',
 ]

vectorphile-0.3.0/vectorphile/autograd/functional.py

@@ -0,0 +1,171 @@
+"""Higher-order autograd utilities: grad, jacobian, hessian, vmap, checkpoint."""
+from __future__ import annotations
+import numpy as np
+from contextlib import contextmanager
+from typing import Callable, List, Optional
+
+_GRAD_ENABLED = True
+
+
+@contextmanager
+def enable_grad():
+    """Force gradient tracking inside a no_grad context."""
+    from vectorphile.autograd.engine import no_grad as _no_grad
+    prev = _no_grad._enabled
+    _no_grad._enabled = True
+    try:
+        yield
+    finally:
+        _no_grad._enabled = prev
+
+
+def grad(f: Callable, x) -> 'Tensor':
+    """Return gradient of scalar f(x) w.r.t. x.
+
+    Args:
+        f: callable Tensor -> scalar Tensor
+        x: input Tensor
+    Returns:
+        gradient Tensor with same shape as x
+    """
+    from vectorphile.core.tensor import Tensor as CoreTensor
+    x.requires_grad = True
+    x.grad = None
+    out = f(x)
+    out.backward(CoreTensor(np.ones_like(out._data)))
+    return CoreTensor(x.grad._data.copy() if x.grad is not None else np.zeros_like(x._data))
+
+
+def jacobian(f: Callable, x) -> np.ndarray:
+    """Full Jacobian of f(x) w.r.t. x via repeated backward passes.
+
+    Args:
+        f: callable Tensor[n] -> Tensor[m]
+        x: 1-D input Tensor
+    Returns:
+        J: np.ndarray shape (m, n)
+    """
+    from vectorphile.core.tensor import Tensor as CoreTensor
+    x_np = x._data.flatten()
+    n = len(x_np)
+    out0_data = f(CoreTensor(x_np))._data.flatten()
+    m = len(out0_data)
+    J = np.zeros((m, n), dtype=np.float32)
+    for i in range(m):
+        xi = CoreTensor(x_np, requires_grad=True)
+        out = f(xi)
+        e = np.zeros(m, dtype=np.float32); e[i] = 1.0
+        out.backward(CoreTensor(e.reshape(out._data.shape)))
+        J[i] = xi.grad._data.flatten() if xi.grad is not None else 0
+    return J
+
+
+def hessian(f: Callable, x) -> np.ndarray:
+    """Hessian of scalar f(x) via central finite differences.
+
+    Args:
+        f: callable Tensor[n] -> scalar Tensor
+        x: 1-D input Tensor
+    Returns:
+        H: np.ndarray shape (n, n)
+    """
+    from vectorphile.core.tensor import Tensor as CoreTensor
+    x_np = x._data.flatten()
+    n = len(x_np)
+    H = np.zeros((n, n), dtype=np.float32)
+    eps = 1e-4
+    for i in range(n):
+        for j in range(n):
+            def _f(v): return f(CoreTensor(v))._data.item()
+            xpp = x_np.copy(); xpp[i] += eps; xpp[j] += eps
+            xpm = x_np.copy(); xpm[i] += eps; xpm[j] -= eps
+            xmp = x_np.copy(); xmp[i] -= eps; xmp[j] += eps
+            xmm = x_np.copy(); xmm[i] -= eps; xmm[j] -= eps
+            H[i,j] = (_f(xpp) - _f(xpm) - _f(xmp) + _f(xmm)) / (4 * eps**2)
+    return H
+
+
+def vmap(f: Callable) -> Callable:
+    """Vectorised map: apply f to each element of a batched Tensor.
+
+    Usage::
+
+        batched = vmap(f)
+        out = batched(x)   # x: (B, ...) applies f to each [i, ...]
+
+    Returns:
+        stacked Tensor of shape (B, *output_shape)
+    """
+    def batched(x):
+        from vectorphile.core.tensor import Tensor as CoreTensor
+        results = [f(CoreTensor(x._data[i])) for i in range(x._data.shape[0])]
+        return CoreTensor(np.stack([r._data for r in results], axis=0))
+    return batched
+
+
+def checkpoint(fn: Callable, *args):
+    """Gradient checkpointing: recompute activations on backward instead of storing.
+
+    Saves memory at the cost of extra compute. The forward pass is run once
+    without grad to get the output; on backward it reruns with grad.
+
+    Args:
+        fn: a callable that takes Tensors and returns a Tensor
+        *args: input Tensors
+    Returns:
+        output Tensor; backward will recompute fn
+    """
+    from vectorphile.core.tensor import Tensor as CoreTensor
+    from vectorphile.autograd.engine import no_grad as _no_grad
+
+    # Forward without grad to get output value
+    saved_enabled = _no_grad._enabled
+    _no_grad._enabled = False
+    with_no_grad_args = [CoreTensor(a._data) for a in args]
+    _no_grad._enabled = False
+    out_data = fn(*with_no_grad_args)._data.copy()
+    _no_grad._enabled = saved_enabled
+
+    out = CoreTensor(out_data, requires_grad=True)
+
+    # On backward: rerun fn with grads
+    class _CheckpointFn:
+        def __init__(self):
+            self.inputs = list(args)
+        def backward(self, grad_out):
+            fresh = [CoreTensor(a._data, requires_grad=a.requires_grad) for a in args]
+            result = fn(*fresh)
+            result.backward(grad_out)
+            return [a.grad for a in fresh]
+
+    out.grad_fn = _CheckpointFn()
+    return out
+
+
+# ── Tensor construction helpers ───────────────────────────────────────────────
+
+def zeros(shape, requires_grad: bool = False):
+    from vectorphile.core.tensor import Tensor as CoreTensor
+    return CoreTensor(np.zeros(shape, dtype=np.float32), requires_grad=requires_grad)
+
+
+def ones(shape, requires_grad: bool = False):
+    from vectorphile.core.tensor import Tensor as CoreTensor
+    return CoreTensor(np.ones(shape, dtype=np.float32), requires_grad=requires_grad)
+
+
+def randn(*shape, requires_grad: bool = False):
+    from vectorphile.core.tensor import Tensor as CoreTensor
+    return CoreTensor(np.random.randn(*shape).astype(np.float32), requires_grad=requires_grad)
+
+
+def cat(tensors: list, axis: int = 0):
+    """Concatenate tensors."""
+    from vectorphile.core.tensor import Tensor as CoreTensor
+    return CoreTensor(np.concatenate([t._data for t in tensors], axis=axis))
+
+
+def stack(tensors: list, axis: int = 0):
+    """Stack tensors along new axis."""
+    from vectorphile.core.tensor import Tensor as CoreTensor
+    return CoreTensor(np.stack([t._data for t in tensors], axis=axis))