Skip to content

README.md

Latest commit

 

History

History
95 lines (68 loc) · 2.56 KB

README.md

File metadata and controls

95 lines (68 loc) · 2.56 KB

Python bindings

pybind11 wrappers over the C++ framework. Install from the repo root:

pip install .
# or editable dev install:
pip install -e ".[dev]"
pytest tests/python -v

Quick examples

Autograd

import numpy as np
import tiramisu as tr

x = tr.from_numpy(np.array([2.0], dtype=np.float32))
x.requires_grad = True
y = tr.add(tr.mul(x, x), tr.mul(x, 3.0))
y.backward()
print(np.asarray(x.grad))  # [7.]

Linear forward + NumPy

import numpy as np
import tiramisu as tr

layer = tr.nn.Linear(784, 10)
x = tr.from_numpy(np.random.randn(2, 784).astype(np.float32))
out = layer.forward(x)
arr = np.asarray(out)  # zero-copy view when contiguous

GPT — explicit causal loss

forward() returns logits. Apply the next-token shift yourself:

import numpy as np
import tiramisu as tr

vocab_size = 65
seq_len = 8
model = tr.nn.GPT(vocab_size=vocab_size, d_model=32, num_layers=1, max_seq_len=seq_len)

input_ids = tr.from_numpy(np.arange(seq_len, dtype=np.float32).reshape(1, seq_len))
logits = model.forward(input_ids)  # (batch, seq, vocab)

logits_np = np.asarray(logits)
targets_np = np.asarray(input_ids)
flat_logits = tr.from_numpy(logits_np[:, :-1, :].reshape(-1, vocab_size))
flat_targets = tr.from_numpy(targets_np[:, 1:].reshape(-1))

loss = tr.nn.cross_entropy_loss(flat_logits, flat_targets)
loss.backward()

Token indices are stored as float32 tensors today (the C++ embedding path reads float indices).

API surface (v1)

Module Symbols
tiramisu Tensor, from_numpy, add, mul, sub, div, relu, gelu, softmax, matmul, sum, mean, reshape, transpose, backward
tiramisu.nn Linear, LayerNorm, GPT, cross_entropy_loss
tiramisu.optim Adam

Tensor

  • Tensor(shape) — float32 CPU tensor
  • from_numpy(arr) — copy from NumPy (float32 or castable)
  • numpy() / np.asarray(t) — zero-copy export when contiguous
  • requires_grad, grad, backward()
  • shape(), numel(), reshape(), slice(), contiguous()

GPT

Keyword args: vocab_size, d_model, num_heads=2, num_layers=2, max_seq_len=256.

Methods: forward(token_ids), parameters(), config().

Build notes

CMake option TIRAMISU_BUILD_PYTHON=ON builds the tiramisu._C extension. The pip build sets this automatically and disables sanitizers/tests for a clean wheel.

cmake -S . -B build -DTIRAMISU_BUILD_PYTHON=ON -DTIRAMISU_BUILD_TESTS=OFF \
  -DTIRAMISU_ENABLE_SANITIZERS=OFF -DCMAKE_BUILD_TYPE=Release
cmake --build build --target _C