mirror of
https://github.com/pytorch/pytorch.git
synced 2025-10-20 12:54:11 +08:00
The big semantic change (and the reason for this port) is that we no longer monkeypatch Tensor with torchdim's special methods. The new algorithm for handling dispatch is that we first land in `__torch_function__` and we see if a special FCD implementation needs to be dispatch to first, and if there is nothing we fallback to the standard level strategy. Because there is no longer C binding equivalent of classes, we've condensed _C.Dim and Dim together, and similar for Tensor. This resulted in some bugs as the Python API is sometimes different from the C API. I've attempted to disambiguate these but there may still be mistakes (many early bugs were due to this problem). Dim and DimEntry are especially painful as Dim must abide by Tensor equality semantics, but is pointer equality in C (DimEntry doesn't have this problem). Another difference between C/Python that is subtle is we no longer get implicit conversions from Dim to DimEntry, this also caused some bugs. Much of the mechanical porting work was done by claude code. I have a separate PR that deletes functorch._C, but it was useful having dim.cpp to point claude at it so I haven't done it in this PR. From a reviewing perspective, I need to re-review that I didn't forget to port anything, some noticeably missing "small" things are patched_dim_method. I am still in progress of carefully doing a side-by-side review of ports; "simplifications" from claude code were also a major source of bugs. There are two major feature gaps in the implementation: - DelayedTensor and dot handling are not implemented yet. This should be reasonably easy, just need to do it. However, for the purposes of sharded propagation it is actually better not to reconstruct matmuls. - Splitting dimensions with an index like `[x, y]` doesn't work. The problem is that `__getitem__` interprets this as advanced indexing and sends the list to torch.tensor to turn into a tensor, instead of being eligible for `__torch_function__`. I think I might need to hard code a special case for this or something? Signed-off-by: Edward Yang <ezyang@meta.com> Pull Request resolved: https://github.com/pytorch/pytorch/pull/160236 Approved by: https://github.com/zdevito, https://github.com/albanD
140 lines
4.8 KiB
Python
140 lines
4.8 KiB
Python
from __future__ import annotations
|
|
|
|
from typing import Any, TYPE_CHECKING
|
|
|
|
import torch
|
|
|
|
from ._dim_entry import DimEntry
|
|
|
|
|
|
if TYPE_CHECKING:
|
|
from . import Dim, Tensor
|
|
|
|
|
|
class EnableAllLayers:
|
|
"""
|
|
RAII-style context manager for enabling functorch vmap layers.
|
|
It manages the creation and cleanup of functorch dynamic layers.
|
|
|
|
This is probably one of the more algorithmically important parts of first
|
|
class dims. Intuitively, FCD can be thought of as another way of using
|
|
vmap, where you don't actually have to vmap at the top level, instead the
|
|
vmaps are implicitly determined by inspecting the bound dimensions on the
|
|
FCD tensors involved in a compute (this is similar to our concept of
|
|
non-lexical modes that we spent a long time talking about years ago). But
|
|
under the hood you still need to actually enable the vmap mode. So once
|
|
FCD has determined all of the dims we are batching over, it needs to
|
|
enable all those layers so functorch can actually apply the batching
|
|
rules. Therefore enable all layers!
|
|
"""
|
|
|
|
levels_start: int
|
|
levels_to_dim: list[Dim]
|
|
|
|
def __init__(self, levels: list[DimEntry]):
|
|
"""
|
|
Initialize and push dynamic layers for all first-class dimensions.
|
|
|
|
Args:
|
|
levels: List of dimension entries to create layers for
|
|
"""
|
|
|
|
from . import Dim
|
|
|
|
self.levels_start = 0
|
|
self.levels_to_dim = []
|
|
|
|
for l in levels:
|
|
if not l.is_positional():
|
|
d = l.dim()
|
|
assert isinstance(d, Dim)
|
|
self.levels_to_dim.append(d)
|
|
|
|
# Sort by level for stable ordering
|
|
self.levels_to_dim.sort(key=lambda d: d._level)
|
|
|
|
def __enter__(self) -> EnableAllLayers: # noqa: PYI034
|
|
# Create functorch dynamic layers
|
|
for i, dim in enumerate(self.levels_to_dim):
|
|
batch_size = dim.size
|
|
level = torch._C._functorch._vmap_increment_nesting(batch_size, "different")
|
|
if i == 0:
|
|
self.levels_start = level
|
|
return self
|
|
|
|
def __exit__(self, exc_type: Any, exc_val: Any, exc_tb: Any) -> None:
|
|
"""Clean up dynamic layers in reverse order."""
|
|
to_remove = self.levels_start + len(self.levels_to_dim) - 1
|
|
for i in range(len(self.levels_to_dim)):
|
|
popped = torch._C._functorch._vmap_decrement_nesting()
|
|
assert popped == to_remove - i, (
|
|
f"Expected layer {to_remove - i}, got {popped}"
|
|
)
|
|
|
|
def from_batched(self, batchedtensor: torch.Tensor, has_device: bool) -> Tensor:
|
|
"""
|
|
Create a Tensor from a batched tensor by unwrapping functorch layers.
|
|
|
|
Args:
|
|
batchedtensor: Batched tensor from functorch operation
|
|
has_device: Whether tensor has device info
|
|
|
|
Returns:
|
|
Tensor with appropriate levels
|
|
"""
|
|
# Create positional levels for base dimensions
|
|
levels: list[DimEntry] = []
|
|
for i in range(-batchedtensor.dim(), 0):
|
|
levels.append(DimEntry(i))
|
|
|
|
tensor = batchedtensor
|
|
|
|
while torch._C._functorch.is_batchedtensor(tensor):
|
|
level = torch._C._functorch.maybe_get_level(tensor)
|
|
assert level is not None
|
|
assert level >= self.levels_start and level < self.levels_start + len(
|
|
self.levels_to_dim
|
|
)
|
|
dim = DimEntry(self.levels_to_dim[level - self.levels_start])
|
|
bdim = torch._C._functorch.maybe_get_bdim(tensor)
|
|
assert bdim is not None
|
|
levels.insert(bdim, dim)
|
|
tensor = torch._C._functorch.get_unwrapped(tensor)
|
|
|
|
from . import Tensor
|
|
|
|
result = Tensor()
|
|
result._tensor = tensor
|
|
result._batchtensor = batchedtensor
|
|
result._has_device = has_device
|
|
result._levels = levels
|
|
return result
|
|
|
|
def inplace_update_layers(
|
|
self, batchtensor: torch.Tensor, levels: list[DimEntry]
|
|
) -> None:
|
|
"""
|
|
Update the levels of a batched tensor in place.
|
|
|
|
This requires the _maybe_unsafe_set_level binding that we'll add to functorch.
|
|
|
|
Args:
|
|
batchtensor: Batched tensor to update
|
|
levels: New levels to set
|
|
"""
|
|
# Check if tensor is batched
|
|
if not torch._C._functorch.is_batchedtensor(batchtensor):
|
|
return
|
|
|
|
impl = batchtensor
|
|
|
|
for i in reversed(range(len(self.levels_to_dim))):
|
|
if impl is None:
|
|
break
|
|
|
|
if any(l == DimEntry(self.levels_to_dim[i]) for l in levels):
|
|
# This is very interesting! The level on batch tensor is
|
|
# meaningless! We set it RIGHT before we go into vmap
|
|
torch._C._functorch._maybe_unsafe_set_level(impl, self.levels_start + i)
|
|
impl = torch._C._functorch.get_unwrapped(impl)
|