Movement

Movement (low level)¤

view ¤

view(shape, *args) -> Self

.view is an alias for .reshape.

Source code in tinygrad/mixin/movement.py

def view(self, shape, *args) -> Self:
  """`.view` is an alias for `.reshape`."""
  return self.reshape(shape, *args)

reshape ¤

reshape(shape, *args) -> Self

Returns a tensor with the same data as the original tensor but with a different shape. shape can be passed as a tuple or as separate arguments.

t = Tensor.arange(6)
print(t.reshape(2, 3).numpy())

[[0 1 2]
 [3 4 5]]

Source code in tinygrad/mixin/movement.py

def reshape(self, shape, *args) -> Self:
  """
  Returns a tensor with the same data as the original tensor but with a different shape.
  `shape` can be passed as a tuple or as separate arguments.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor.arange(6)
  print(t.reshape(2, 3).numpy())
  ```
  """
  # resolve None and args
  new_shape = tuple([s if s is not None else self.shape[i] for i, s in enumerate(argfix(shape, *args))])
  # resolve -1
  if (c := new_shape.count(-1)) > 1:
    raise RuntimeError(f"only one dimension can be inferred using -1, getting {new_shape}")
  if c:
    new_shape = tuple([-prod(self.shape) // prod(new_shape) if s == -1 else s for s in new_shape])
  if prod(self.shape) != prod(new_shape):
    raise ValueError(f"size mismatch, can't reshape ({self.shape}) -> ({new_shape})")
  ret = self._mop(Ops.RESHAPE, arg=new_shape)
  return self if ret.shape == self.shape else ret

expand ¤

expand(shape, *args) -> Self

Returns a tensor that is expanded to the shape that is specified. Expand can also increase the number of dimensions that a tensor has.

Passing a -1 or None to a dimension means that its size will not be changed.

t = Tensor([1, 2, 3])
print(t.expand(4, -1).numpy())

[[1 2 3]
 [1 2 3]
 [1 2 3]
 [1 2 3]]

Source code in tinygrad/mixin/movement.py

def expand(self, shape, *args) -> Self:
  """
  Returns a tensor that is expanded to the shape that is specified.
  Expand can also increase the number of dimensions that a tensor has.

  Passing a `-1` or `None` to a dimension means that its size will not be changed.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor([1, 2, 3])
  print(t.expand(4, -1).numpy())
  ```
  """
  new_shape = tuple(from_ if to == -1 or to is None else to for from_, to in zip(*(_align_left(self.shape, argfix(shape, *args)))))
  return self._broadcast_to(new_shape)

permute ¤

permute(order, *args) -> Self

Returns a tensor that is a permutation of the original tensor. The new tensor has the same data as the original tensor but with the dimensions permuted according to the order specified. order can be passed as a tuple or as separate arguments.

t = Tensor.empty(2, 3, 5)
print(t.shape)

(2, 3, 5)

print(t.permute(2, 0, 1).shape)

(5, 2, 3)

Source code in tinygrad/mixin/movement.py

def permute(self, order, *args) -> Self:
  """
  Returns a tensor that is a permutation of the original tensor.
  The new tensor has the same data as the original tensor but with the dimensions permuted according to the order specified.
  `order` can be passed as a tuple or as separate arguments.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor.empty(2, 3, 5)
  print(t.shape)
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(t.permute(2, 0, 1).shape)
  ```
  """
  order_arg = tuple(self._resolve_dim(x) for x in argfix(order, *args))
  if sorted(order_arg) != list(range(self.ndim)):
    raise RuntimeError(f"order is not a valid permutation, getting {order_arg}")
  return self._mop(Ops.PERMUTE, arg=order_arg) if order_arg != tuple(range(self.ndim)) else self

flip ¤

flip(axis, *args) -> Self

Returns a tensor that reverses the order of the original tensor along given axis. axis can be passed as a tuple or as separate arguments.

t = Tensor.arange(6).reshape(2, 3)
print(t.numpy())

[[0 1 2]
 [3 4 5]]

print(t.flip(0).numpy())

[[3 4 5]
 [0 1 2]]

print(t.flip((0, 1)).numpy())

[[5 4 3]
 [2 1 0]]

Source code in tinygrad/mixin/movement.py

def flip(self, axis, *args) -> Self:
  """
  Returns a tensor that reverses the order of the original tensor along given `axis`.
  `axis` can be passed as a tuple or as separate arguments.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor.arange(6).reshape(2, 3)
  print(t.numpy())
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(t.flip(0).numpy())
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(t.flip((0, 1)).numpy())
  ```
  """
  axis_arg = tuple(self._resolve_dim(x) for x in argfix(axis, *args))
  assert all(not isinstance(x, bool) and x >= 0 and x < self.ndim for x in axis_arg), f"flip args must be axis ints {axis_arg}"
  if len(axis_arg) != len(dedup(axis_arg)):
    raise RuntimeError(f"dim can appear at most once, getting {axis_arg}")
  flip_arg = tuple([i in axis_arg for i in range(len(self.shape))])
  return self._mop(Ops.FLIP, arg=flip_arg) if any(flip_arg) else self

shrink ¤

shrink(arg: tuple[tuple[sint, sint] | None, ...]) -> Self

Returns a tensor that shrinks the each axis based on input arg. arg must have the same length as self.ndim. For each axis, it can be None, which means no shrink, or a tuple (start, end) that works the same as Python slice.

t = Tensor.arange(9).reshape(3, 3)
print(t.numpy())

[[0 1 2]
 [3 4 5]
 [6 7 8]]

print(t.shrink(((None, (1, 3)))).numpy())

[[1 2]
 [4 5]
 [7 8]]

print(t.shrink((((0, 2), (0, 2)))).numpy())

[[0 1]
 [3 4]]

Source code in tinygrad/mixin/movement.py

def shrink(self, arg: tuple[tuple[sint, sint] | None, ...]) -> Self:
  """
  Returns a tensor that shrinks the each axis based on input arg.
  `arg` must have the same length as `self.ndim`.
  For each axis, it can be `None`, which means no shrink, or a tuple `(start, end)` that works the same as Python slice.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor.arange(9).reshape(3, 3)
  print(t.numpy())
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(t.shrink(((None, (1, 3)))).numpy())
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(t.shrink((((0, 2), (0, 2)))).numpy())
  ```
  """
  if self.ndim != len(arg):
    raise ValueError(f"{self.ndim=} != {len(arg)=}")
  ret = self._mop(Ops.SHRINK, arg=[x if x is not None else (0, s) for x, s in zip(arg, self.shape)])
  return self if ret.shape == self.shape else ret

pad ¤

pad(
    padding: (
        Sequence[sint] | Sequence[tuple[sint, sint] | None]
    ),
    mode: str = "constant",
    value: float = 0.0,
) -> Tensor

Returns a tensor with padding applied based on the input padding.

padding supports two padding structures:

Flat padding: (padding_left, padding_right, padding_top, padding_bottom, ...)
- This structure matches PyTorch's pad.
- padding length must be even.
Group padding: (..., (padding_top, padding_bottom), (padding_left, padding_right))
- This structure matches pad for JAX, NumPy, TensorFlow, and others.
- For each axis, padding can be None, meaning no padding, or a tuple (start, end).
- padding must have the same length as self.ndim.

Padding values can be negative, resulting in dimension shrinks that work similarly to Python negative slices. Padding modes is selected with mode which supports constant, reflect and replicate.

t = Tensor.arange(9).reshape(1, 1, 3, 3)
print(t.numpy())

[[[[0 1 2]
   [3 4 5]
   [6 7 8]]]]

print(t.pad((1, 2, 0, -1)).numpy())

[[[[0 0 1 2 0 0]
   [0 3 4 5 0 0]]]]

print(t.pad(((None, None, (0, -1), (1, 2)))).numpy())

[[[[0 0 1 2 0 0]
   [0 3 4 5 0 0]]]]

print(t.pad((1, 2, 0, -1), value=-float('inf')).numpy())

[[[[-inf   0.   1.   2. -inf -inf]
   [-inf   3.   4.   5. -inf -inf]]]]

Source code in tinygrad/tensor.py

def pad(self, padding:Sequence[sint]|Sequence[tuple[sint, sint]|None], mode:str="constant", value:float=0.0) -> Tensor:
  """
  Returns a tensor with padding applied based on the input `padding`.

  `padding` supports two padding structures:

  1. Flat padding: `(padding_left, padding_right, padding_top, padding_bottom, ...)`
      - This structure matches PyTorch's pad.
      - `padding` length must be even.

  2. Group padding: `(..., (padding_top, padding_bottom), (padding_left, padding_right))`
      - This structure matches pad for JAX, NumPy, TensorFlow, and others.
      - For each axis, padding can be `None`, meaning no padding, or a tuple `(start, end)`.
      - `padding` must have the same length as `self.ndim`.

  Padding values can be negative, resulting in dimension shrinks that work similarly to Python negative slices.
  Padding modes is selected with `mode` which supports `constant`, `reflect` and `replicate`.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor.arange(9).reshape(1, 1, 3, 3)
  print(t.numpy())
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(t.pad((1, 2, 0, -1)).numpy())
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(t.pad(((None, None, (0, -1), (1, 2)))).numpy())
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(t.pad((1, 2, 0, -1), value=-float('inf')).numpy())
  ```
  """
  # normalize to grouped format
  if all(isinstance(p, (int,UOp)) for p in padding):
    if len(padding)%2 != 0: raise ValueError("Flat padding must have even number of pads")
    pX = _flat_to_grouped(tuple(cast(Sequence[sint], padding)) + (0,0)*(self.ndim - len(padding)//2))
  else: pX = tuple((0,0) if p is None else p for p in cast(Sequence[tuple[sint, sint]|None], padding))
  if len(pX) != self.ndim: raise ValueError(f"padding length is improper, {padding=} {self.ndim=}")
  # dispatch
  if mode == "constant": return self._pad_constant(pX, value)
  assert all_int(self.shape), f"does not support symbolic shape {self.shape}"
  if mode == "circular": return self._pad_circular(pX)
  if mode in {"reflect", "replicate"}: return self._pad_reflect_replicate(pX, mode)
  raise NotImplementedError(f"{mode=} is not supported")

Movement (high level)¤

getitem ¤

__getitem__(indices) -> Tensor

Retrieves a sub-tensor using indexing.

Examples:

t = Tensor.arange(12).reshape(3, 4)
print(t.numpy())

[[ 0  1  2  3]
 [ 4  5  6  7]
 [ 8  9 10 11]]

Int Indexing: Select an element or sub-tensor using integers for each dimension.
```
print(t[1, 2].numpy())
```
```
6
```
Slice Indexing: Select a range of elements using slice notation (start:end:stride).
```
print(t[0:2, ::2].numpy())
```
```
[[0 2]
 [4 6]]
```
Tensor Indexing: Use another tensor as indices for advanced indexing. Using tuple or list here also works.
```
print(t[Tensor([2, 0, 1]), Tensor([1, 2, 3])].numpy())
```
```
[9 2 7]
```
None Indexing: Add a new dimension to the tensor.
```
print(t[:, None].shape)
```
```
(3, 1, 4)
```

Note

Out-of-bounds indexing results in a value of 0.

t = Tensor([1, 2, 3])
print(t[Tensor([4, 3, 2])].numpy())

[0 0 3]

Source code in tinygrad/tensor.py

def __getitem__(self, indices) -> Tensor:
  """
  Retrieves a sub-tensor using indexing.

  Supported Index Types: `int | slice | Tensor | None | list | tuple | Ellipsis`

  Examples:
  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor.arange(12).reshape(3, 4)
  print(t.numpy())
  ```

  - Int Indexing: Select an element or sub-tensor using integers for each dimension.
    ```python exec="true" source="above" session="tensor" result="python"
    print(t[1, 2].numpy())
    ```

  - Slice Indexing: Select a range of elements using slice notation (`start:end:stride`).
    ```python exec="true" source="above" session="tensor" result="python"
    print(t[0:2, ::2].numpy())
    ```

  - Tensor Indexing: Use another tensor as indices for advanced indexing. Using `tuple` or `list` here also works.
    ```python exec="true" source="above" session="tensor" result="python"
    print(t[Tensor([2, 0, 1]), Tensor([1, 2, 3])].numpy())
    ```

  - `None` Indexing: Add a new dimension to the tensor.
    ```python exec="true" source="above" session="tensor" result="python"
    print(t[:, None].shape)
    ```

  NOTE: Out-of-bounds indexing results in a value of `0`.
  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor([1, 2, 3])
  print(t[Tensor([4, 3, 2])].numpy())
  ```
  """
  return self._getitem(indices)

gather ¤

gather(dim: int, index: Tensor) -> Tensor

Gathers values along an axis specified by dim.

t = Tensor([[1, 2], [3, 4]])
print(t.numpy())

[[1 2]
 [3 4]]

print(t.gather(1, Tensor([[0, 0], [1, 0]])).numpy())

[[1 1]
 [4 3]]

Source code in tinygrad/tensor.py

def gather(self:Tensor, dim:int, index:Tensor) -> Tensor:
  """
  Gathers values along an axis specified by `dim`.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor([[1, 2], [3, 4]])
  print(t.numpy())
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(t.gather(1, Tensor([[0, 0], [1, 0]])).numpy())
  ```
  """
  assert index.ndim == self.ndim, f"self.ndim must equal index.ndim, {self.ndim=}, {index.ndim=}"
  dim = self._resolve_dim(dim)
  assert all(s >= i for d,(s,i) in enumerate(zip(self.shape, index.shape)) if d != dim), "requires self.shape[d] >= index.shape[d] for all d != dim"
  index = index.to(self.device)
  x = self.shrink_to(tuple(i if d != dim else None for d,i in enumerate(index.shape))).unsqueeze(-1).transpose(-1, dim)
  return (index.unsqueeze(-1)._one_hot_along_dim(self.shape[dim]).where(x, 0)).sum(-1, dtype=self.dtype)

cat ¤

cat(*args: Self, dim: int = 0) -> Self

Concatenates self with other tensors in args along an axis specified by dim. All tensors must have the same shape except in the concatenating dimension.

t0, t1, t2 = Tensor([[1, 2]]), Tensor([[3, 4]]), Tensor([[5, 6]])
print(t0.cat(t1, t2, dim=0).numpy())

[[1 2]
 [3 4]
 [5 6]]

print(t0.cat(t1, t2, dim=1).numpy())

[[1 2 3 4 5 6]]

Source code in tinygrad/mixin/__init__.py

def cat(self, *args:Self, dim:int=0) -> Self:
  """
  Concatenates self with other tensors in `args` along an axis specified by `dim`.
  All tensors must have the same shape except in the concatenating dimension.

  ```python exec="true" source="above" session="tensor" result="python"
  t0, t1, t2 = Tensor([[1, 2]]), Tensor([[3, 4]]), Tensor([[5, 6]])
  print(t0.cat(t1, t2, dim=0).numpy())
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(t0.cat(t1, t2, dim=1).numpy())
  ```
  """
  dim = self._resolve_dim(dim)
  for arg in args: assert arg.ndim==self.ndim and all(ti==ai for i,(ti,ai) in enumerate(zip(self.shape, arg.shape)) if i!=dim)
  tensors = [self, *args]
  dim_cumsum = list(itertools.accumulate([t.shape[dim] for t in tensors], initial=0))
  padded = [t.pad(tuple((dim_cumsum[i], dim_cumsum[-1]-dim_cumsum[i+1]) if j==dim else None for j in range(t.ndim))) for i,t in enumerate(tensors)]
  return padded[0].usum(*padded[1:])

stack ¤

stack(*args: Self, dim: int = 0) -> Self

Concatenates self with other tensors in args along a new dimension specified by dim.

t0, t1, t2 = Tensor([1, 2]), Tensor([3, 4]), Tensor([5, 6])
print(t0.stack(t1, t2, dim=0).numpy())

[[1 2]
 [3 4]
 [5 6]]

print(t0.stack(t1, t2, dim=1).numpy())

[[1 3 5]
 [2 4 6]]

Source code in tinygrad/mixin/__init__.py

def stack(self, *args:Self, dim:int=0) -> Self:
  """
  Concatenates self with other tensors in `args` along a new dimension specified by `dim`.

  ```python exec="true" source="above" session="tensor" result="python"
  t0, t1, t2 = Tensor([1, 2]), Tensor([3, 4]), Tensor([5, 6])
  print(t0.stack(t1, t2, dim=0).numpy())
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(t0.stack(t1, t2, dim=1).numpy())
  ```
  """
  # checks for shapes and number of dimensions delegated to cat
  unsqueezed = [t.unsqueeze(dim) for t in argfix(self, *args)]
  return unsqueezed[0].cat(*unsqueezed[1:], dim=dim)

repeat ¤

repeat(repeats, *args) -> Self

Repeats tensor number of times along each dimension specified by repeats. repeats can be passed as a tuple or as separate arguments.

t = Tensor([1, 2, 3])
print(t.repeat(4, 2).numpy())

[[1 2 3 1 2 3]
 [1 2 3 1 2 3]
 [1 2 3 1 2 3]
 [1 2 3 1 2 3]]

print(t.repeat(4, 2, 1).shape)

(4, 2, 3)

Source code in tinygrad/mixin/movement.py

def repeat(self, repeats, *args) -> Self:
  """
  Repeats tensor number of times along each dimension specified by `repeats`.
  `repeats` can be passed as a tuple or as separate arguments.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor([1, 2, 3])
  print(t.repeat(4, 2).numpy())
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(t.repeat(4, 2, 1).shape)
  ```
  """
  repeats = argfix(repeats, *args)
  base_shape = _align_left(self.shape, repeats)[0]
  unsqueezed_shape = flatten([[s] if r == 1 else [1, s] for r, s in zip(repeats, base_shape)])
  expanded_shape = flatten([[s] if r == 1 else [r, s] for r, s in zip(repeats, base_shape)])
  final_shape = [r * s for r, s in zip(repeats, base_shape)]
  return self.reshape(unsqueezed_shape).expand(expanded_shape).reshape(final_shape)

repeat_interleave ¤

repeat_interleave(
    repeats: int, dim: int | None = None
) -> Self

Repeats elements of a tensor.

t = Tensor([1, 2, 3])
print(t.repeat_interleave(2).numpy())

[1 1 2 2 3 3]

Source code in tinygrad/mixin/movement.py

def repeat_interleave(self, repeats: int, dim: int | None = None) -> Self:
  """
  Repeats elements of a tensor.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor([1, 2, 3])
  print(t.repeat_interleave(2).numpy())
  ```
  """
  x, dim = (self.flatten(), 0) if dim is None else (self, self._resolve_dim(dim))
  shp = x.shape
  x = x.reshape(*shp[: dim + 1], 1, *shp[dim + 1 :])
  x = x.expand(*shp[: dim + 1], repeats, *shp[dim + 1 :])
  x = x.reshape(*shp[:dim], shp[dim] * repeats, *shp[dim + 1 :])
  return x

split ¤

split(
    sizes: int | Sequence[int], dim: int = 0
) -> tuple[Self, ...]

Splits the tensor into chunks along the dimension specified by dim. If sizes is an integer, it splits into equally sized chunks if possible, otherwise the last chunk will be smaller. If sizes is a list, it splits into len(sizes) chunks with size in dim according to size.

t = Tensor.arange(10).reshape(5, 2)
print(t.numpy())

[[0 1]
 [2 3]
 [4 5]
 [6 7]
 [8 9]]

split = t.split(2)
print("\n".join([repr(x.numpy()) for x in split]))

array([[0, 1],
       [2, 3]], dtype=int32)
array([[4, 5],
       [6, 7]], dtype=int32)
array([[8, 9]], dtype=int32)

split = t.split([1, 4])
print("\n".join([repr(x.numpy()) for x in split]))

array([[0, 1]], dtype=int32)
array([[2, 3],
       [4, 5],
       [6, 7],
       [8, 9]], dtype=int32)

Source code in tinygrad/mixin/movement.py

def split(self, sizes:int|Sequence[int], dim:int=0) -> tuple[Self, ...]:
  """
  Splits the tensor into chunks along the dimension specified by `dim`.
  If `sizes` is an integer, it splits into equally sized chunks if possible, otherwise the last chunk will be smaller.
  If `sizes` is a list, it splits into `len(sizes)` chunks with size in `dim` according to `size`.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor.arange(10).reshape(5, 2)
  print(t.numpy())
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  split = t.split(2)
  print("\\n".join([repr(x.numpy()) for x in split]))
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  split = t.split([1, 4])
  print("\\n".join([repr(x.numpy()) for x in split]))
  ```
  """
  dim = self._resolve_dim(dim)
  dim_sz = self.shape[dim]
  assert isinstance(dim_sz, int), f"does not support symbolic shape in split dimension {dim}: {self.shape}"
  if isinstance(sizes, int): sizes = [min(sizes, dim_sz-i) for i in range(0, max(1, dim_sz), max(1, sizes))]
  assert sum(sizes) == dim_sz, f"expect sizes to sum exactly to {dim_sz}, but got {sum(sizes)}"
  return tuple(self.shrink(tuple((sum(sizes[:i]), sum(sizes[:i+1])) if j == dim else None for j in range(self.ndim))) for i in range(len(sizes)))

chunk ¤

chunk(chunks: int, dim: int = 0) -> list[Self]

Splits the tensor into chunks number of chunks along the dimension dim. If the tensor size along dim is not divisible by chunks, all returned chunks will be the same size except the last one. The function may return fewer than the specified number of chunks.

chunked = Tensor.arange(11).chunk(6)
print("\n".join([repr(x.numpy()) for x in chunked]))

array([0, 1], dtype=int32)
array([2, 3], dtype=int32)
array([4, 5], dtype=int32)
array([6, 7], dtype=int32)
array([8, 9], dtype=int32)
array([10], dtype=int32)

chunked = Tensor.arange(12).chunk(6)
print("\n".join([repr(x.numpy()) for x in chunked]))

array([0, 1], dtype=int32)
array([2, 3], dtype=int32)
array([4, 5], dtype=int32)
array([6, 7], dtype=int32)
array([8, 9], dtype=int32)
array([10, 11], dtype=int32)

chunked = Tensor.arange(13).chunk(6)
print("\n".join([repr(x.numpy()) for x in chunked]))

array([0, 1, 2], dtype=int32)
array([3, 4, 5], dtype=int32)
array([6, 7, 8], dtype=int32)
array([ 9, 10, 11], dtype=int32)
array([12], dtype=int32)

Source code in tinygrad/mixin/movement.py

def chunk(self, chunks:int, dim:int=0) -> list[Self]:
  """
  Splits the tensor into `chunks` number of chunks along the dimension `dim`.
  If the tensor size along `dim` is not divisible by `chunks`, all returned chunks will be the same size except the last one.
  The function may return fewer than the specified number of chunks.

  ```python exec="true" source="above" session="tensor" result="python"
  chunked = Tensor.arange(11).chunk(6)
  print("\\n".join([repr(x.numpy()) for x in chunked]))
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  chunked = Tensor.arange(12).chunk(6)
  print("\\n".join([repr(x.numpy()) for x in chunked]))
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  chunked = Tensor.arange(13).chunk(6)
  print("\\n".join([repr(x.numpy()) for x in chunked]))
  ```
  """
  dim = self._resolve_dim(dim)
  dim_sz = self.shape[dim]
  assert isinstance(dim_sz, int), f"does not support symbolic shape in split dimension {dim}: {self.shape}"
  assert chunks > 0, f"expect chunks to be greater than 0, got: {chunks}"
  return list(self.split(ceildiv(dim_sz, chunks) if dim_sz else [0]*chunks, dim=dim))

unfold ¤

unfold(dim: int, size, step: int) -> Self

Unfolds the tensor along dimension dim into overlapping windows.

Each window has length size and begins every step elements of self. Returns the input tensor with dimension dim replaced by dims (n_windows, size) where n_windows = (self.shape[dim] - size) // step + 1.

unfolded = Tensor.arange(8).unfold(0,2,2)
print("\n".join([repr(x.numpy()) for x in unfolded]))

array([0, 1], dtype=int32)
array([2, 3], dtype=int32)
array([4, 5], dtype=int32)
array([6, 7], dtype=int32)

unfolded = Tensor.arange(27).reshape(3,3,3).unfold(-1,2,3)
print("\n".join([repr(x.numpy()) for x in unfolded]))

array([[[0, 1]],

       [[3, 4]],

       [[6, 7]]], dtype=int32)
array([[[ 9, 10]],

       [[12, 13]],

       [[15, 16]]], dtype=int32)
array([[[18, 19]],

       [[21, 22]],

       [[24, 25]]], dtype=int32)

Source code in tinygrad/mixin/movement.py

def unfold(self, dim:int, size, step:int) -> Self:
  """
  Unfolds the tensor along dimension `dim` into overlapping windows.

  Each window has length `size` and begins every `step` elements of `self`.
  Returns the input tensor with dimension `dim` replaced by dims `(n_windows, size)`
  where `n_windows = (self.shape[dim] - size) // step + 1`.

  ```python exec="true" source="above" session="tensor" result="python"
  unfolded = Tensor.arange(8).unfold(0,2,2)
  print("\\n".join([repr(x.numpy()) for x in unfolded]))
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  unfolded = Tensor.arange(27).reshape(3,3,3).unfold(-1,2,3)
  print("\\n".join([repr(x.numpy()) for x in unfolded]))
  ```
  """
  if size < 0: raise RuntimeError(f'size must be >= 0 but got {size=}')
  if step <= 0: raise RuntimeError(f'step must be > 0 but got {step=}')
  if size > self.shape[dim]: raise RuntimeError(f'maximum size for tensor at dimension {dim} is {self.shape[dim]} but size is {size}')
  dim = self._resolve_dim(dim)
  perm_to_last = tuple(i for i in range(self.ndim) if i != dim) + (dim,)
  return self.permute(perm_to_last)._pool((size,), step).permute(argsort(perm_to_last) + (self.ndim,))

meshgrid ¤

meshgrid(*args, indexing: str = 'ij') -> tuple[Self, ...]

Generates coordinate matrices from coordinate vectors. Input tensors can be scalars or 1D tensors.

indexing determines how the output grids are aligned. ij indexing follows matrix-style indexing and xy indexing follows Cartesian-style indexing.

x, y = Tensor([1, 2, 3]), Tensor([4, 5, 6])
grid_x, grid_y = x.meshgrid(y)
print(grid_x.numpy())
print(grid_y.numpy())

[[1 1 1]
 [2 2 2]
 [3 3 3]]
[[4 5 6]
 [4 5 6]
 [4 5 6]]

grid_x, grid_y = x.meshgrid(y, indexing="xy")
print(grid_x.numpy())
print(grid_y.numpy())

[[1 2 3]
 [1 2 3]
 [1 2 3]]
[[4 4 4]
 [5 5 5]
 [6 6 6]]

Source code in tinygrad/mixin/movement.py

def meshgrid(self, *args, indexing:str="ij") -> tuple[Self, ...]:
  """
  Generates coordinate matrices from coordinate vectors.
  Input tensors can be scalars or 1D tensors.

  `indexing` determines how the output grids are aligned.
  `ij` indexing follows matrix-style indexing and `xy` indexing follows Cartesian-style indexing.

  ```python exec="true" source="above" session="tensor" result="python"
  x, y = Tensor([1, 2, 3]), Tensor([4, 5, 6])
  grid_x, grid_y = x.meshgrid(y)
  print(grid_x.numpy())
  print(grid_y.numpy())
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  grid_x, grid_y = x.meshgrid(y, indexing="xy")
  print(grid_x.numpy())
  print(grid_y.numpy())
  ```
  """
  if indexing not in ("ij", "xy"): raise RuntimeError(f'indexing must be in ("ij", "xy"), got {indexing}')
  if len(tensors:=(self, *args)) == 1: return tensors
  basis = tuple(range(len(tensors))) if indexing == "ij" else (1, 0) + tuple(range(2, len(tensors)))
  tensors = tuple(t.reshape((-1,) + (1,)*(len(args) - i)) for i,t in zip(basis, tensors))
  output_shape = _broadcast_shape(*(t.shape for t in tensors))
  return tuple(t._broadcast_to(output_shape) for t in tensors)

squeeze ¤

squeeze(dim: int | None = None) -> Self

Returns a tensor with specified dimensions of input of size 1 removed. If dim is not specified, all dimensions with size 1 are removed.

t = Tensor.zeros(2, 1, 2, 1, 2)
print(t.squeeze().shape)

(2, 2, 2)

print(t.squeeze(0).shape)

(2, 1, 2, 1, 2)

print(t.squeeze(1).shape)

(2, 2, 1, 2)

Source code in tinygrad/mixin/movement.py

def squeeze(self, dim: int | None = None) -> Self:
  """
  Returns a tensor with specified dimensions of input of size 1 removed.
  If `dim` is not specified, all dimensions with size 1 are removed.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor.zeros(2, 1, 2, 1, 2)
  print(t.squeeze().shape)
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(t.squeeze(0).shape)
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(t.squeeze(1).shape)
  ```
  """
  if dim is None:
    return self.reshape(tuple(dim for dim in self.shape if dim != 1))
  dim = self._resolve_dim(dim)
  return self if not self.ndim or self.shape[dim] != 1 else self.reshape(self.shape[:dim] + self.shape[dim + 1 :])

unsqueeze ¤

unsqueeze(dim: int) -> Self

Returns a tensor with a new dimension of size 1 inserted at the specified dim.

t = Tensor([1, 2, 3, 4])
print(t.unsqueeze(0).numpy())

[[1 2 3 4]]

print(t.unsqueeze(1).numpy())

[[1]
 [2]
 [3]
 [4]]

Source code in tinygrad/mixin/movement.py

def unsqueeze(self, dim: int) -> Self:
  """
  Returns a tensor with a new dimension of size 1 inserted at the specified `dim`.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor([1, 2, 3, 4])
  print(t.unsqueeze(0).numpy())
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(t.unsqueeze(1).numpy())
  ```
  """
  dim = self._resolve_dim(dim, extra=True)
  return self.reshape(self.shape[:dim] + (1,) + self.shape[dim:])

T `property` ¤

T: Self

.T is an alias for .transpose().

transpose ¤

transpose(dim0=1, dim1=0) -> Self

Returns a tensor that is a transposed version of the original tensor. The given dimensions dim0 and dim1 are swapped.

t = Tensor.arange(6).reshape(2, 3)
print(t.numpy())

[[0 1 2]
 [3 4 5]]

print(t.transpose(0, 1).numpy())

[[0 3]
 [1 4]
 [2 5]]

Source code in tinygrad/mixin/movement.py

def transpose(self, dim0=1, dim1=0) -> Self:
  """
  Returns a tensor that is a transposed version of the original tensor.
  The given dimensions `dim0` and `dim1` are swapped.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor.arange(6).reshape(2, 3)
  print(t.numpy())
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(t.transpose(0, 1).numpy())
  ```
  """
  order = list(range(self.ndim))
  order[dim0], order[dim1] = order[dim1], order[dim0]
  return self.permute(order)

flatten ¤

flatten(start_dim=0, end_dim=-1) -> Self

Flattens the tensor by reshaping it into a one-dimensional tensor. If start_dim or end_dim are passed, only dimensions starting with start_dim and ending with end_dim are flattened.

t = Tensor.arange(8).reshape(2, 2, 2)
print(t.flatten().numpy())

[0 1 2 3 4 5 6 7]

print(t.flatten(start_dim=1).numpy())

[[0 1 2 3]
 [4 5 6 7]]

Source code in tinygrad/mixin/movement.py

def flatten(self, start_dim=0, end_dim=-1) -> Self:
  """
  Flattens the tensor by reshaping it into a one-dimensional tensor.
  If `start_dim` or `end_dim` are passed, only dimensions starting with `start_dim` and ending with `end_dim` are flattened.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor.arange(8).reshape(2, 2, 2)
  print(t.flatten().numpy())
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(t.flatten(start_dim=1).numpy())
  ```
  """
  start_dim, end_dim = self._resolve_dim(start_dim), self._resolve_dim(end_dim)
  return self.reshape(self.shape[:start_dim] + (prod(self.shape[start_dim : end_dim + 1]),) + self.shape[end_dim + 1 :])

unflatten ¤

unflatten(dim: int, sizes: tuple[int, ...]) -> Self

Unflattens dimension dim of the tensor into multiple dimensions specified by sizes. Tensor.flatten() is the inverse of this function.

print(Tensor.ones(3, 4, 1).unflatten(1, (2, 2)).shape)

(3, 2, 2, 1)

print(Tensor.ones(3, 4, 1).unflatten(1, (-1, 2)).shape)

(3, 2, 2, 1)

print(Tensor.ones(5, 12, 3).unflatten(-2, (2, 2, 3, 1, 1)).shape)

(5, 2, 2, 3, 1, 1, 3)

Source code in tinygrad/mixin/movement.py

def unflatten(self, dim: int, sizes: tuple[int, ...]) -> Self:
  """
  Unflattens dimension `dim` of the tensor into multiple dimensions specified by `sizes`. `Tensor.flatten()` is the inverse of this function.

  ```python exec="true" source="above" session="tensor" result="python"
  print(Tensor.ones(3, 4, 1).unflatten(1, (2, 2)).shape)
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(Tensor.ones(3, 4, 1).unflatten(1, (-1, 2)).shape)
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(Tensor.ones(5, 12, 3).unflatten(-2, (2, 2, 3, 1, 1)).shape)
  ```
  """
  dim = self._resolve_dim(dim)
  return self.reshape(self.shape[:dim] + sizes + self.shape[dim + 1 :])

diag ¤

diag() -> Self

Returns a 2-D square tensor with the elements of input as the main diagonal.

print(Tensor([1, 2, 3]).diag().numpy())

[[1 0 0]
 [0 2 0]
 [0 0 3]]

Source code in tinygrad/mixin/movement.py

def diag(self) -> Self:
  """
  Returns a 2-D square tensor with the elements of input as the main diagonal.

  ```python exec="true" source="above" session="tensor" result="python"
  print(Tensor([1, 2, 3]).diag().numpy())
  ```
  """
  if self.ndim != 1: raise ValueError(f"expect input to be 1-D, getting {self.ndim}-D")
  return self.unsqueeze(-1).pad_to((None, 1+(n:=self.shape[0]))).flatten().shrink_to((n*n,)).reshape(n,n)

diagonal ¤

diagonal(
    offset: int = 0, dim1: int = 0, dim2: int = 1
) -> Self

Returns a view of the diagonal elements with respect to dim1 and dim2. offset controls which diagonal: 0 is main, positive is above, negative is below.

t = Tensor.arange(9).reshape(3, 3)
print(t.numpy())

[[0 1 2]
 [3 4 5]
 [6 7 8]]

print(t.diagonal().numpy())

[0 4 8]

print(t.diagonal(offset=1).numpy())

[1 5]

Source code in tinygrad/mixin/movement.py

def diagonal(self, offset:int=0, dim1:int=0, dim2:int=1) -> Self:
  """
  Returns a view of the diagonal elements with respect to `dim1` and `dim2`.
  `offset` controls which diagonal: 0 is main, positive is above, negative is below.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor.arange(9).reshape(3, 3)
  print(t.numpy())
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(t.diagonal().numpy())
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(t.diagonal(offset=1).numpy())
  ```
  """
  if (dim1:=self._resolve_dim(dim1)) == (dim2:=self._resolve_dim(dim2)): raise RuntimeError("dim1 and dim2 cannot be the same dimension")
  x = self.permute(*[i for i in range(self.ndim) if i != dim1 and i != dim2], dim1, dim2)
  if offset >= 0: x = x.shrink(tuple(None for _ in x.shape[:-1]) + ((offset, x.shape[-1]),))
  else: x = x.shrink(tuple(None for _ in x.shape[:-2]) + ((-offset, x.shape[-2]), None))
  if (d := min(int(x.shape[-2]), int(x.shape[-1]))) <= 0: return x.reshape(*x.shape[:-2], 0)
  nones, x = tuple(None for _ in x.shape[:-2]), x.shrink_to(tuple(None for _ in x.shape[:-2]) + (d, d))
  return x.flatten(-2).pad_to(nones+(d*(d+1),)).unflatten(-1, (d, d+1)).shrink_to(nones+(None, 1)).squeeze(-1)

roll ¤

roll(
    shifts: int | tuple[int, ...],
    dims: int | tuple[int, ...] | None = None,
) -> Self

Rolls the tensor along specified dimension(s). The rolling operation is circular, meaning that elements that go beyond the edge are wrapped around to the beginning of the dimension.

t = Tensor.arange(4)
print(t.roll(shifts=1, dims=0).numpy())

[3 0 1 2]

print(t.roll(shifts=-1, dims=0).numpy())

[1 2 3 0]

Source code in tinygrad/mixin/movement.py

def roll(self, shifts:int|tuple[int, ...], dims:int|tuple[int, ...]|None=None) -> Self:
  """
  Rolls the tensor along specified dimension(s).
  The rolling operation is circular, meaning that elements that go beyond the edge are wrapped around to the beginning of the dimension.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor.arange(4)
  print(t.roll(shifts=1, dims=0).numpy())
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(t.roll(shifts=-1, dims=0).numpy())
  ```
  """
  if dims is None: return self.flatten().roll(shifts, 0).reshape(self.shape)
  dims, shifts = tuple(self._resolve_dim(d) for d in make_tuple(dims, 1)), make_tuple(shifts, 1)
  if len(dims) != len(shifts): raise RuntimeError(f"{len(dims)=} != {len(shifts)=}")
  shrink_arg: list[tuple[sint, sint]|None] = [None] * self.ndim
  for d, s in zip(dims, shifts): shrink_arg[d] = (delta:=self.shape[d]-s%self.shape[d], delta+self.shape[d])
  return self.repeat(*tuple(2 if i in dims else 1 for i in range(self.ndim))).shrink(tuple(shrink_arg))

rearrange ¤

rearrange(formula: str, **sizes) -> Self

Rearranges input according to formula

See: https://einops.rocks/api/rearrange/

x = Tensor([[1, 2], [3, 4]])
print(Tensor.rearrange(x, "batch channel -> (batch channel)").numpy())

[1 2 3 4]

Source code in tinygrad/mixin/movement.py

def rearrange(self, formula: str, **sizes) -> Self:
  """
  Rearranges input according to formula

  See: https://einops.rocks/api/rearrange/

  ```python exec="true" source="above" session="tensor" result="python"
  x = Tensor([[1, 2], [3, 4]])
  print(Tensor.rearrange(x, "batch channel -> (batch channel)").numpy())
  ```
  """

  def parse_side(s: str) -> tuple[list[str], list[tuple[int, int]]]:
    """Parse one side of formula into (axis_names, dims) where dims are (start, end) index pairs for parens."""
    tokens = f" {s} ".replace("…", "...").replace("(", " ( ").replace(")", " ) ").replace(" ", "  ").replace(" 1 ", " ( ) ").split()
    lparens, rparens = [i for i, tok in enumerate(tokens) if tok == "("], [i for i, tok in enumerate(tokens) if tok == ")"]
    pairs = list(zip(lparens, rparens))
    assert len(lparens) == len(rparens) and sorted(flatten(pairs)) == flatten(pairs), "bracket mismatch"
    return [tok for tok in tokens if tok not in ("(", ")")], [(lp - 2*i, rp - 1 - 2*i) for i, (lp, rp) in enumerate(pairs)]

  assert formula.count("->") == 1, 'need exactly one "->" in formula'
  (lhs, unflatten_dims), (rhs, flatten_dims) = map(parse_side, formula.split("->"))

  for name in sizes: assert name in lhs, f"axis {name} is not used in transform"
  assert sorted(lhs) == sorted(rhs) and len(lhs) == len(set(lhs)), f"name mismatch in {formula}"
  for name in lhs+rhs: assert name == "..." or (name.isidentifier() and "_" not in (name[0], name[-1])), f"invalid axis name {name}"
  assert "..." not in flatten([lhs[s:e] for s, e in unflatten_dims]), f"cannot have collapsed ellipsis (...) in lhs of {formula}"
  assert lhs.count("...") <= 1, f"too many ellipses in {formula}"

  # resolve ellipsis
  if "..." in lhs:
    ell_len = len(self.shape) - len(lhs) + 1 + sum(e - s - 1 for s, e in unflatten_dims)
    lhs, rhs = map(lambda l: l[:(i := l.index("..."))] + [f"...{j}" for j in range(ell_len)] + l[i + 1:] if "..." in l else l, (lhs, rhs))
    def newdims(side, s, e): return (s + (ell_len - 1 if "...0" in side[:s] else 0), e + (ell_len - 1 if "...0" in side[:e] else 0))
    unflatten_dims, flatten_dims = [newdims(lhs, s, e) for s, e in unflatten_dims], [newdims(rhs, s, e) for s, e in flatten_dims]

  # unflatten -> permute -> flatten
  t = self
  for start, end in unflatten_dims: t = t.unflatten(start, tuple(sizes.get(lhs[i], -1) for i in range(start, end)))
  for i, name in enumerate(lhs):
    if name in sizes: assert sizes[name] == t.shape[i], f"size provided for dimension {name} incorrect"
  t = t.permute([lhs.index(name) for name in rhs])
  for start, end in reversed(flatten_dims): t = t.flatten(start, end - 1) if start < end else t.unsqueeze(start)
  return t

Movement

Movement (low level)¤

view ¤

reshape ¤

expand ¤

permute ¤

flip ¤

shrink ¤

pad ¤

Movement (high level)¤

__getitem__ ¤

gather ¤

cat ¤

stack ¤

repeat ¤

repeat_interleave ¤

split ¤

chunk ¤

unfold ¤

meshgrid ¤

squeeze ¤

unsqueeze ¤

T property ¤

transpose ¤

flatten ¤

unflatten ¤

diag ¤

diagonal ¤

roll ¤

rearrange ¤

getitem ¤

T `property` ¤