Use the examples to see tinyfin without external frameworks. Run from repo root after make libtinyfin.so.

• mnist_mlp.py — synthetic MNIST-like MLP (CPU, numpy only).
  python examples/mnist_mlp.py

• cifar_cnn.py — synthetic CIFAR-like conv + pool + linear classifier (CPU).
  python examples/cifar_cnn.py

• transformer_tiny.py — transformer-inspired token-mixing/FFN block on random data (no external deps).
  python examples/transformer_tiny.py

• transformer_block_tiny.py — tiny transformer-style block (embedding + FFN + residual).
  python examples/transformer_block_tiny.py

• text_gen_tiny.py — tiny character-level text generation on a short string.
  python examples/text_gen_tiny.py

• rnn_tiny.py — minimal LSTM sequence example with a tiny training loop.
  python examples/rnn_tiny.py

• cuda_matmul.py — CUDA matmul demo if built with ENABLE_CUDA=1 (falls back to CPU).
  TINYFIN_BACKEND=cuda python examples/cuda_matmul.py

• perf_profile.py — throughput/GFLOP/s profiler (matmul/conv2d/elem).
  python examples/perf_profile.py cpu matmul 512 512 512 20
python examples/perf_profile.py cpu conv2d 16 3 32 32 8 3 20
python examples/perf_profile.py cpu elem 1048576 50

• graph_export.py — exports a tiny autograd graph to DOT.
  python examples/graph_export.py

• save_load.py — saves and reloads a tiny linear model + optimizer state.
  python examples/save_load.py

• backend_mnist_cnn.py — synthetic MNIST-like CNN with backend toggle.
  TINYFIN_BACKEND=cpu|cuda|opengl|vulkan python examples/backend_mnist_cnn.py

Notes:

• For CUDA, build with ENABLE_CUDA=1 and set TINYFIN_BACKEND=cuda.
• For BLAS matmul backend, build with ENABLE_BLAS=1 and set TINYFIN_BACKEND=blas.