Properties

Basic

shape property

shape: Tuple[sint, ...]

dtype property

dtype: DType

device property

device: Union[str, Tuple[str, ...]]

ndim property

ndim: int

Returns the number of dimensions in the tensor.

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

2

numel

numel() -> sint

Returns the total number of elements in the tensor.

t = Tensor([[[1, 2], [3, 4]], [[5, 6], [7, 8]]])
print(t.numel())

8

Source code in tinygrad/tensor.py

def numel(self) -> sint:
  """
  Returns the total number of elements in the tensor.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor([[[1, 2], [3, 4]], [[5, 6], [7, 8]]])
  print(t.numel())
  ```
  """
  return prod(self.shape)

element_size

element_size() -> int

Returns the size in bytes of an individual element in the tensor.

t = Tensor([5], dtype=dtypes.int16)
print(t.element_size())

2

Source code in tinygrad/tensor.py

def element_size(self) -> int:
  """
  Returns the size in bytes of an individual element in the tensor.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor([5], dtype=dtypes.int16)
  print(t.element_size())
  ```
  """
  return self.dtype.itemsize

nbytes

nbytes() -> int

Returns the total number of bytes of all elements in the tensor.

t = Tensor([8, 9], dtype=dtypes.float)
print(t.nbytes())

8

Source code in tinygrad/tensor.py

def nbytes(self) -> int:
  """
  Returns the total number of bytes of all elements in the tensor.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor([8, 9], dtype=dtypes.float)
  print(t.nbytes())
  ```
  """
  return self.numel() * self.element_size()

is_floating_point

is_floating_point() -> bool

Returns True if the tensor contains floating point types, i.e. is one of dtype.float64, dtype.float32, dtype.float16, dtype.bfloat16.

t = Tensor([8, 9], dtype=dtypes.float32)
print(t.is_floating_point())

True

Source code in tinygrad/tensor.py

def is_floating_point(self) -> bool:
  """
  Returns `True` if the tensor contains floating point types, i.e. is one of `dtype.float64`, `dtype.float32`,
  `dtype.float16`, `dtype.bfloat16`.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor([8, 9], dtype=dtypes.float32)
  print(t.is_floating_point())
  ```
  """
  return dtypes.is_float(self.dtype)

size

size(
    dim: Optional[int] = None,
) -> Union[sint, Tuple[sint, ...]]

Return the size of the tensor. If dim is specified, return the length along dimension dim. Otherwise return the shape of the tensor.

t = Tensor([[4, 5, 6], [7, 8, 9]])
print(t.size())

(2, 3)

print(t.size(dim=1))

3

Source code in tinygrad/tensor.py

def size(self, dim:Optional[int]=None) -> Union[sint, Tuple[sint, ...]]:
  """
  Return the size of the tensor. If `dim` is specified, return the length along dimension `dim`. Otherwise return the shape of the tensor.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor([[4, 5, 6], [7, 8, 9]])
  print(t.size())
  ```
  ```python exec="true" source="above" session="tensor" result="python"
  print(t.size(dim=1))
  ```
  """
  return self.shape if dim is None else self.shape[dim]

Data Access

data

data() -> memoryview

Returns the data of this tensor as a memoryview.

t = Tensor([1, 2, 3, 4])
print(np.frombuffer(t.data(), dtype=np.int32))

[1 2 3 4]

Source code in tinygrad/tensor.py

def data(self) -> memoryview:
  """
  Returns the data of this tensor as a memoryview.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor([1, 2, 3, 4])
  print(np.frombuffer(t.data(), dtype=np.int32))
  ```
  """
  assert self.dtype.base.fmt is not None, f"no fmt dtype for {self.dtype.base}"
  assert all_int(self.shape), f"no data if shape is symbolic, {self.shape=}"
  if TYPE_CHECKING or sys.version_info < (3, 12): assert self.dtype.base.fmt != "e"
  return self._data().cast(self.dtype.base.fmt) if 0 in self.shape else self._data().cast(self.dtype.base.fmt, self.shape)

item

item() -> ConstType

Returns the value of this tensor as a standard Python number.

t = Tensor(42)
print(t.item())

42

Source code in tinygrad/tensor.py

def item(self) -> ConstType:
  """
  Returns the value of this tensor as a standard Python number.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor(42)
  print(t.item())
  ```
  """
  assert self.numel() == 1, "must have one element for item"
  return self.data()[(0,) * len(self.shape)]

tolist

tolist() -> Union[Sequence[ConstType], ConstType]

Returns the value of this tensor as a nested list.

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

[1, 2, 3, 4]

Source code in tinygrad/tensor.py

def tolist(self) -> Union[Sequence[ConstType], ConstType]:
  """
  Returns the value of this tensor as a nested list.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor([1, 2, 3, 4])
  print(t.tolist())
  ```
  """
  return self.data().tolist()

numpy

numpy() -> 'np.ndarray'

Returns the value of this tensor as a numpy.ndarray.

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

array([1, 2, 3, 4], dtype=int32)

Source code in tinygrad/tensor.py

def numpy(self) -> 'np.ndarray':  # type: ignore [name-defined] # noqa: F821
  """
  Returns the value of this tensor as a `numpy.ndarray`.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor([1, 2, 3, 4])
  print(repr(t.numpy()))
  ```
  """
  import numpy as np
  if self.dtype.base == dtypes.bfloat16: return self.float().numpy()
  assert _to_np_dtype(self.dtype.base) is not None, f"no np dtype for {self.dtype.base}"
  assert all_int(self.shape), f"no data if shape is symbolic, {self.shape=}"
  return np.frombuffer(self._data(), dtype=_to_np_dtype(self.dtype.base)).reshape(self.shape)

tinygrad ops

schedule_with_vars

schedule_with_vars(
    *lst: Tensor,
) -> Tuple[List[ScheduleItem], Dict[Variable, int]]

Creates the schedule needed to realize these Tensor(s), with Variables.

Note

A Tensor can only be scheduled once.

Source code in tinygrad/tensor.py

def schedule_with_vars(self, *lst:Tensor) -> Tuple[List[ScheduleItem], Dict[Variable, int]]:
  """
  Creates the schedule needed to realize these Tensor(s), with Variables.

  NOTE: A Tensor can only be scheduled once.
  """
  schedule, var_vals = create_schedule_with_vars(flatten([x.lazydata.lbs for x in (self,)+lst]))
  return memory_planner(schedule), var_vals

schedule

schedule(*lst: Tensor) -> List[ScheduleItem]

Creates the schedule needed to realize these Tensor(s).

Source code in tinygrad/tensor.py

def schedule(self, *lst:Tensor) -> List[ScheduleItem]:
  """Creates the schedule needed to realize these Tensor(s)."""
  schedule, var_vals = self.schedule_with_vars(*lst)
  assert len(var_vals) == 0
  return schedule

realize

realize(*lst: Tensor, do_update_stats=True) -> Tensor

Triggers the computation needed to create these Tensor(s).

Source code in tinygrad/tensor.py

def realize(self, *lst:Tensor, do_update_stats=True) -> Tensor:
  """Triggers the computation needed to create these Tensor(s)."""
  run_schedule(*self.schedule_with_vars(*lst), do_update_stats=do_update_stats)
  return self

replace

replace(x: Tensor) -> Tensor

Replaces the data of this tensor with the data of another tensor. Only the shape of the tensors must match.

Source code in tinygrad/tensor.py

def replace(self, x:Tensor) -> Tensor:
  """
  Replaces the data of this tensor with the data of another tensor. Only the shape of the tensors must match.
  """
  # used for replacing a Tensor with a new version of it (potentially with a different device and dtype)
  assert not x.requires_grad and getattr(self, '_ctx', None) is None
  assert self.shape == x.shape, f"replace shape mismatch {self.shape} != {x.shape}"
  self.lazydata = x.lazydata
  return self

assign

assign(x) -> Tensor

Source code in tinygrad/tensor.py

def assign(self, x) -> Tensor:
  # TODO: this is a hack for writing to DISK. remove with working assign
  if isinstance(self.device, str) and self.device.startswith("DISK"):
    if x.__class__ is not Tensor: x = Tensor(x, device="CLANG", dtype=self.dtype)
    self.contiguous().realize().lazydata.base.realized.copyin(x._data())
    return self
  if x.__class__ is not Tensor: x = Tensor(x, device=self.device, dtype=self.dtype)
  if DEBUG >= 4: print(f"assign {self.lazydata} <- {x.lazydata}")
  if self.lazydata is x.lazydata: return self  # a self assign is a NOOP
  # NOTE: we allow cross device assign
  assert self.shape == x.shape, f"assign shape mismatch {self.shape} != {x.shape}"
  assert self.device == x.device, f"assign device mismatch {self.device} != {x.device}"
  assert self.dtype == x.dtype, f"assign dtype mismatch {self.dtype} != {x.dtype}"
  assert not isinstance(self.lazydata, MultiLazyBuffer) or self.lazydata.axis == x.lazydata.axis, "axis must match on MultiLazyBuffer"
  assert not x.requires_grad  # self requires_grad is okay?
  if not self.lazydata.is_realized: return self.replace(x)
  self.lazydata = self.lazydata.assign(x.lazydata)
  return self

detach

detach() -> Tensor

Returns a new tensor with the same data as this tensor, but detached from the autograd graph.

Source code in tinygrad/tensor.py

def detach(self) -> Tensor:
  """
  Returns a new tensor with the same data as this tensor, but detached from the autograd graph.
  """
  return Tensor(self.lazydata, device=self.device, requires_grad=False)

to

to(device: Optional[Union[str, Tuple[str, ...]]]) -> Tensor

Moves the tensor to the given device.

Source code in tinygrad/tensor.py

def to(self, device:Optional[Union[str, Tuple[str, ...]]]) -> Tensor:
  """
  Moves the tensor to the given device.
  """
  device = tuple(Device.canonicalize(x) for x in device) if isinstance(device, (tuple, list)) else Device.canonicalize(device)
  if device == self.device: return self
  if not isinstance(device, str): return self.shard(device)
  ret = Tensor(self.lazydata, device, requires_grad=self.requires_grad)
  if self.grad is not None: ret.grad = self.grad.to(device)
  if hasattr(self, '_ctx'): ret._ctx = self._ctx
  return ret

to_

to_(device: Optional[Union[str, Tuple[str, ...]]])

Moves the tensor to the given device in place.

Source code in tinygrad/tensor.py

def to_(self, device:Optional[Union[str, Tuple[str, ...]]]):
  """
  Moves the tensor to the given device in place.
  """
  real = self.to(device)
  # TODO: is this assign?
  if self.grad is not None and real.grad is not None: self.grad.lazydata = real.grad.lazydata
  self.lazydata = real.lazydata

shard

shard(
    devices: Tuple[str, ...],
    axis: Optional[int] = None,
    splits: Optional[Tuple[int, ...]] = None,
) -> Tensor

Shards the tensor across the given devices. Optionally specify which axis to shard on, and how to split it across devices.

t = Tensor.empty(2, 3)
print(t.shard((t.device, t.device), axis=1, splits=(2, 1)).lazydata)

<MLB self.axis=1 self.real=[True, True] 
CLANG ShapeTracker(views=(View(shape=(2, 2), strides=(2, 1), offset=0, mask=None, contiguous=True),))
CLANG ShapeTracker(views=(View(shape=(2, 1), strides=(1, 0), offset=0, mask=None, contiguous=True),))>

Source code in tinygrad/tensor.py

def shard(self, devices:Tuple[str, ...], axis:Optional[int]=None, splits:Optional[Tuple[int, ...]]=None) -> Tensor:
  """
  Shards the tensor across the given devices. Optionally specify which axis to shard on, and how to split it across devices.

  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor.empty(2, 3)
  print(t.shard((t.device, t.device), axis=1, splits=(2, 1)).lazydata)
  ```

  """
  assert isinstance(self.lazydata, LazyBuffer), "can't shard a MultiLazyBuffer"
  devices, bounds = tuple(Device.canonicalize(x) for x in devices), None
  if axis is not None:
    if axis < 0: axis += len(self.shape)
    if splits is None:
      if not isinstance(total:=self.shape[axis], int): raise RuntimeError(f"cannot shard symbolic shape {self.shape=}, {axis=}")
      sz = ceildiv(total, len(devices))
      splits = tuple([max(0, min(sz, total - sz*i)) for i in range(len(devices))])
    assert sum(splits) == self.shape[axis], "specified splits do not sum up to axis shape"
    boundaries = tuple(itertools.accumulate(splits))
    bounds = tuple(zip((0,) + boundaries, boundaries))
  return Tensor(MultiLazyBuffer.from_sharded(self.lazydata, devices, axis, bounds), device=devices, requires_grad=self.requires_grad)

shard_

shard_(
    devices: Tuple[str, ...],
    axis: Optional[int] = None,
    splits: Optional[Tuple[int, ...]] = None,
)

Shards the tensor across the given devices in place.

Source code in tinygrad/tensor.py

def shard_(self, devices:Tuple[str, ...], axis:Optional[int]=None, splits:Optional[Tuple[int, ...]]=None):
  """
  Shards the tensor across the given devices in place.
  """
  self.lazydata = self.shard(devices, axis, splits).lazydata
  return self

contiguous

contiguous()

Returns a contiguous tensor.

Source code in tinygrad/tensor.py

def contiguous(self):
  """
  Returns a contiguous tensor.
  """
  return F.Contiguous.apply(self)

contiguous_backward

contiguous_backward()

Inserts a contiguous operation in the backward pass.

Source code in tinygrad/tensor.py

def contiguous_backward(self):
  """
  Inserts a contiguous operation in the backward pass.
  """
  return F.ContiguousBackward.apply(self)

backward

backward(
    gradient: Optional[Tensor] = None,
    retain_graph: bool = False,
) -> Tensor

Propagates the gradient of a tensor backwards through the computation graph. If the 'gradient' argument is not provided, the tensor must be a scalar, and the gradient is implicitly set to 1.0. If 'retain_graph' is false, the graph used to compute the grads will be freed. Otherwise, it will be kept. Keeping it can increase memory usage.

t = Tensor([1.0, 2.0, 3.0, 4.0], requires_grad=True)
t.sum().backward()
print(t.grad.numpy())

[1. 1. 1. 1.]

Source code in tinygrad/tensor.py

def backward(self, gradient:Optional[Tensor]=None, retain_graph:bool=False) -> Tensor:
  """
  Propagates the gradient of a tensor backwards through the computation graph.
  If the 'gradient' argument is not provided, the tensor must be a scalar, and the gradient is implicitly set to 1.0.
  If 'retain_graph' is false, the graph used to compute the grads will be freed. Otherwise, it will be kept. Keeping it can increase memory usage.
  ```python exec="true" source="above" session="tensor" result="python"
  t = Tensor([1.0, 2.0, 3.0, 4.0], requires_grad=True)
  t.sum().backward()
  print(t.grad.numpy())
  ```
  """
  toposorted = self._deepwalk()
  if gradient is None:
    assert self.shape == tuple(), "when no gradient is provided, backward must be called on a scalar tensor"
    # fill in the first grad with one. don't use Tensor.ones because we don't need contiguous
    # this is "implicit gradient creation"
    gradient = Tensor(1.0, dtype=self.dtype, device=self.device, requires_grad=False)

  assert self.shape == gradient.shape, f"grad shape must match tensor shape, {gradient.shape!r} != {self.shape!r}"
  self.grad = gradient
  for t0 in reversed(toposorted):
    if t0.grad is None: raise RuntimeError(f"tensor {t0} has no grad")
    token = _METADATA.set(dataclasses.replace(md, backward=True) if (md := t0._ctx.metadata) is not None else None)
    grads = t0._ctx.backward(t0.grad.lazydata)
    _METADATA.reset(token)
    grads = [Tensor(g, device=self.device, requires_grad=False) if g is not None else None
      for g in ([grads] if len(t0._ctx.parents) == 1 else grads)]
    for t, g in zip(t0._ctx.parents, grads):
      if g is not None and t.requires_grad:
        assert g.shape == t.shape, f"grad shape must match tensor shape, {g.shape!r} != {t.shape!r}"
        t.grad = g if t.grad is None else (t.grad + g)
    if not retain_graph: del t0._ctx
  return self