rasptorch

rasptorch is an experimental deep learning library inspired by PyTorch, with a specific goal: make training and running neural networks practical on the Raspberry Pi 5 while taking advantage of its GPU.

The project has two main parts:

• A small, NumPy-backed autograd engine and nn module that runs on the Raspberry Pi CPU.
• An experimental Vulkan-based backend, wired through a device API (Tensor(..., device="cpu"|"gpu").to("gpu")), meant to offload core tensor operations (elementwise ops, matmul, activations, reductions, etc.) to the Pi 5's GPU via Vulkan compute.

The Vulkan backend is implemented with real Vulkan compute shaders (GLSL compiled to SPIR-V). It supports a small but useful set of kernels (see rasptorch/shaders/ for the authoritative list).

High-level highlights:

• Elementwise ops: +, *, -, neg, relu, gelu, silu, leaky_relu, elu (plus scalar variants)
• Matmul: @ (tiled/shared-memory shader)
• Tensor helpers: indexing/slicing via tensor[...], unsqueeze, squeeze, permute, transpose, flatten, max, min, argmax, argmin, cat, stack, split, chunk
• Reductions: sum, mean (global and axis-based on CPU; global on GPU)
• Common broadcast forms: (N,M) + (M,) and (N,M) * (M,)
• Losses: cross_entropy, binary_cross_entropy, binary_cross_entropy_with_logits, nll_loss, smooth_l1_loss, label_smoothing_cross_entropy
• Optimizers and training utilities: SGD, Adam, AdamW, RMSProp, LR schedulers, gradient clipping, regularization helpers
• NN essentials: GPU row-wise softmax / log_softmax, 2D LayerNorm, BatchNorm1d, BatchNorm2d, Embedding, MultiheadAttention, MaxPool2d, AvgPool2d, and GRU

Performance notes:

• The fastest path is compute-only: keep tensors on GPU and avoid per-iteration .numpy() / readbacks.
• Fusing ops and reusing output buffers can make the Vulkan path faster than NumPy for certain workloads on Raspberry Pi 5.

See main.py for a simple training example and gpu_demo.py for a focused correctness + benchmark suite for the Vulkan backend.

Demos

• Essentials demo (softmax/log_softmax, LayerNorm, Dropout, no_grad, detach):
  • CPU: uv run essentials_demo.py --device cpu
  • GPU-autograd (Vulkan): uv run essentials_demo.py --device gpu

Note: the Vulkan-backed GPU path requires working Vulkan drivers and glslc (shader compiler) on your PATH. Low-level backend helpers can still fall back to NumPy in some environments, but main.py --device gpu is expected to run with Vulkan (glslc available) and fails clearly if it is not.

Modes

There are three execution modes exposed via main.py --device ...:

• cpu: NumPy autograd engine (PyTorch-like, runs on CPU)
• gpu: explicit Vulkan training path (forward + backward + SGD via purpose-built kernels)
• gpu-autograd: experimental GPU autograd (builds a graph on GPU for a growing but still incomplete set of ops)

Quickstart

• Use a virtual environment for best results (e.g. .venv, .venv + uv, .venv + poetry).

Installation

From PyPI (CPU-only):

• pip install rasptorch

GPU (Pi 5 Vulkan):

• pip install "rasptorch[gpu]"

Optional (for saving/loading .pth via real torch.save/torch.load):

• pip install "rasptorch[torch]"

Dev/test:

• pip install -e ".[dev]"

Notes for GPU mode:

• Requires working Vulkan drivers on your system.
• Requires glslc (shader compiler) available on PATH.

Quick GPU validation:

• uv run gpu_demo.py --smoke-only
  • Initializes Vulkan strictly and runs fast correctness checks for core kernels.
  • If this fails, uv run main.py --device gpu will also fail.

Quick model saving check:

• uv run main.py --device cpu --epochs 1 --save model.pth
  • If torch is installed: python -c "import torch; print(torch.load('model.pth').keys())"
  • If not: python -c "import pickle; print(pickle.load(open('model.pth','rb')).keys())"

For local development from this repo:

• pip install -e .

GPU Training (Vulkan)

There are currently two “modes” of training in this repo:

• CPU autograd training (PyTorch-like): uses the NumPy-backed autograd engine.
• Vulkan GPU training (explicit kernels): runs forward + backward + SGD updates on GPU using purpose-built compute shaders.

The Vulkan training path lives in rasptorch/gpu_training.py and currently supports a 2-layer MLP:

Linear -> ReLU -> Linear with MSE loss and SGD.

The general gpu-autograd path supports substantially more model-building pieces than the explicit gpu trainer, including adaptive optimizers, additional activations, LayerNorm, BatchNorm, Embedding, and MultiheadAttention.

Run it via:

• uv run main.py --device gpu --epochs 50 --batch-size 32 --lr 0.1

Saving weights (PyTorch-style .pth):

• uv run main.py --device gpu --epochs 50 --save model.pth
• If torch is installed, this is a real torch.save(...) file loadable via torch.load("model.pth").
• If torch is not installed, rasptorch falls back to writing a pickle payload (same keys, not torch.load compatible).

GPU Autograd (WIP)

There is now an experimental gpu-autograd mode that enables loss.backward() even when the model and activations live on GPU, for a limited set of ops.

Run it via:

• uv run main.py --device gpu-autograd --epochs 50 --batch-size 32 --lr 0.1

Currently supported (GPU) in autograd:

• +, *, - (scalar and tensor forms), @ (matmul)
• scalar ops: tensor + s, tensor * s, tensor / s, plus s + tensor, s * tensor, s - tensor
• neg, relu, gelu, silu, leaky_relu, elu, sum, mean, T (2D transpose)
• tensor shape/join helpers: unsqueeze, squeeze, flatten, permute (common tensors up to 4D), transpose(dim0, dim1), cat, stack, split, chunk
• functional.softmax / functional.log_softmax (2D row-wise, dim=-1/1)
• nn.LayerNorm (2D inputs, 1D normalized_shape; eps=1e-5 stays on GPU, other eps values fall back to CPU)
• nn.BatchNorm1d, nn.BatchNorm2d, nn.Embedding, nn.MultiheadAttention
• Linear backward (GPU grads for weight/bias)
• SGD.step() updates GPU parameters in-place (SGD + optional momentum/weight decay)
• Adam.step(), AdamW.step(), RMSProp.step() update GPU parameters in-place
• functional.cross_entropy(logits, target_onehot) (softmax cross-entropy, mean reduction)

Also available on the CPU autograd path:

• GRU backward
• Tensor.__getitem__ / slicing
• axis-based sum(axis=...) / mean(axis=...)
• max(axis=...) / min(axis=...) with autograd
• argmax(axis=...) / argmin(axis=...)
• nn.MaxPool2d / nn.AvgPool2d

Also available across the library:

• LR schedulers: StepLR, MultiStepLR, ExponentialLR, CosineAnnealingLR, ReduceLROnPlateau, WarmupScheduler
• Initialization helpers: kaiming_*, xavier_*, orthogonal_, uniform_, normal_, zeros_, ones_, constant_
• Gradient utilities: clip_grad_norm_, clip_grad_value_, l1_regularization, l2_regularization, total_variation_loss
• AMP surface: rasptorch.amp.autocast() and rasptorch.amp.GradScaler

Tip: rasptorch.no_grad() exists (like PyTorch) to disable graph building during evaluation.

Training Loop Utilities

There is now a small, reusable training loop helper in rasptorch.train that provides PyTorch-like epoch logs (loss, accuracy/metrics, throughput) for any model.

Key pieces:

• rasptorch.train.fit(...): train loop with optional validation
• rasptorch.train.Accuracy(): top-1 classification accuracy
• rasptorch.train.classification_target_one_hot(C, device=...): converts integer labels -> one-hot

Example (classifier):

from rasptorch import functional as F
from rasptorch.train import fit, Accuracy, classification_target_one_hot
from rasptorch.optim import SGD

model = ...
opt = SGD(model.parameters(), lr=0.01, momentum=0.9, weight_decay=1e-4)

fit(
    model,
    opt,
    train_loader,
    loss_fn=F.cross_entropy,
    device="gpu",
    epochs=10,
    val_loader=val_loader,
    target_transform=classification_target_one_hot(num_classes=10, device="gpu"),
    metrics=[Accuracy()],
)

Notes:

• Metrics like accuracy call .numpy() on logits, which triggers a GPU readback.
• rasptorch.no_grad() exists; evaluation can avoid building graphs.
• mse_loss is now implemented purely via tensor ops ((pred-target)^2 + mean()), so the loss tensor itself is on GPU in gpu-autograd mode; training code typically reads it back via .numpy() for logging.
• Parameters and gradients stay on GPU; loss is read back to CPU for logging.
• uv run main.py --device gpu now requires Vulkan. If Vulkan init or shader compilation fails, it raises a clear error instead of silently falling back.
• Broadcasting is still limited; common 2D + 1D row-vector forms like (N,M) + (M,) and (N,M) * (M,) are supported.

Additional APIs

Optimization:

• rasptorch.optim: SGD, Adam, AdamW, RMSProp
• rasptorch.optim_sched: StepLR, MultiStepLR, ExponentialLR, CosineAnnealingLR, ReduceLROnPlateau, WarmupScheduler

Initialization:

• rasptorch.init: kaiming_uniform_, kaiming_normal_, xavier_uniform_, xavier_normal_, orthogonal_, constant_, zeros_, ones_, uniform_, normal_

Regularization and gradient helpers:

• rasptorch.utils: clip_grad_norm_, clip_grad_value_, l1_regularization, l2_regularization, total_variation_loss

Tensor helpers:

• Tensor.unsqueeze(), Tensor.squeeze(), Tensor.permute(), Tensor.transpose(), Tensor.flatten()
• Tensor.split(), Tensor.chunk()
• rasptorch.cat(...), rasptorch.stack(...)

GPU notes for tensor helpers:

• unsqueeze, squeeze, and flatten are view-based on GPU
• cat, stack, split, and chunk now use Vulkan device-to-device buffer copies
• permute / general transpose(dim0, dim1) are Vulkan-native for common tensors up to 4D

More modules:

• rasptorch.nn: BatchNorm1d, BatchNorm2d, Embedding, MultiheadAttention, GRU, MaxPool2d, AvgPool2d, GELU, SiLU, LeakyReLU, ELU

Mixed precision surface:

• rasptorch.amp.autocast()
• rasptorch.amp.GradScaler
• Tensor.half() / Tensor.float()

Benchmarks

gpu_demo.py prints timing stats (min/p50/p95/mean/std) for:

• CPU (NumPy)
• GPU compute+readback (includes .numpy() every iteration)
• GPU compute-only (no per-iteration readback)
• GPU fused compute-only and no-alloc variants (preallocated output buffers)

If you want the GPU to win, focus on the compute-only + fused/no-alloc numbers.

Current Limitations

• GPU autograd is still incomplete. Core MLP/classification paths are covered, but full PyTorch-like operator coverage is not there yet.
• Some newer APIs use CPU-backed math internally when no dedicated Vulkan/autograd kernel exists yet. The public API works, but not every path is fully GPU-native.
• The newer GPU-native tensor helper coverage is strongest for practical tensors up to 4D; generic higher-rank permutation is not on a dedicated Vulkan path yet.
• GRU autograd is currently CPU-backed; there is no dedicated Vulkan GRU autograd path yet.
• The mixed-precision API surface exists, but true fp16 Vulkan storage/compute kernels are not implemented yet. autocast() and GradScaler are currently preparatory/experimental.
• GPU reductions still focus on the common paths: global sum() / mean() are GPU-native, while axis-based reductions currently fall back to CPU.
• PyTorch integration is experimental: rasptorch.torch_bridge currently supports a small inference subset (Conv2d, Linear, ReLU, BatchNorm2d, MaxPool2d, Sigmoid, Tanh, GELU, Dropout) and may copy tensors CPU<->GPU.

Development & Tests

• pytest runs CPU tests by default.
• The backend smoke test runs everywhere:
  • With Vulkan available, it exercises real GPU kernels.
  • Without Vulkan, it exercises the NumPy fallback path.
• For a strict Vulkan-only check, run uv run gpu_demo.py --smoke-only.

Publishing (maintainers)

Build:

• python -m pip install -U build twine
• python -m build

Upload PyPI:

• python -m twine upload dist/*