Three fixes:
1. When doing t[u0] +=1 if u0 is unbacked we could allocate a new unbacked symbol during the the indexing of t[u0] (when we fake trace setitem), namely because meta_select does allocate a new unbacked symbol for the storage offset when we do not know if u0>=0 or u0<0. but the output size/stride of setitem(), does not depend on that new symbol. it's self consumed in setitem so we shall ignore it.
2. Also when we trace through generalized_scatter the applications of the views could allocate unbacked symints
but those do not effect final output, we also shall ignore them.
3.Before accessing strides in lowering we shall materialize.
Address https://github.com/pytorch/pytorch/issues/114293 and https://github.com/pytorch/pytorch/issues/131911
Pull Request resolved: https://github.com/pytorch/pytorch/pull/164341
Approved by: https://github.com/bobrenjc93
Summary:
The implementation adds the ability to:
Set custom metadata strings that will be attached to all subsequent allocations
Clear or change the metadata at any point
View the metadata in memory snapshots via _dump_snapshot()
Test Plan: Added test in test_cuda.py and check manually in snapshot to see that metadata was added.
Differential Revision: D84654933
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165490
Approved by: https://github.com/yushangdi
- During op dispatch local tensor is supposed to collect rng state from CPU and CUDA
devices so that it can be reset before execution of the op for each such that ops
with randomness produces the same result for all ranks (note that we are planning a
separate change to add support of per rank rng state). Previously we relied on
op input arguments to deduce which devices to get rng state from. Which doesn't work
for factory functions such torch.randn. Hence this changes switches to uncondionally
collecting rng state from all devices.
- Fixing per rank specific computations in _MaskedPartial and Shard placements discovered
during test enablement.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165716
Approved by: https://github.com/ezyang
Summary: `is_backend_available` takes in a string and expects it to only be backend, if its given a composite (device:backend) string, it fails.
Reviewed By: prashrock
Differential Revision: D81886736
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165061
Approved by: https://github.com/H-Huang
Eager AC/SAC reapplies the mutations (like global dict mutations) in the backward during the recomputation of forward. torch.compile has no easy way to reapply python mutations in the backward. But many users might be ok to skip reapplication of side effects in the backward. They can set this config flag to accept this eager and compile divergence.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165775
Approved by: https://github.com/zou3519
ghstack dependencies: #165734
https://github.com/pytorch/pytorch/pull/164820 introduced a bug that `_StridedShard` will call parent class `Shard`'s `split_tensor` method, thus results in incorrect data locality. (I think @ezyang spotted this issue, but we have no test to capture this)
Meanwhile, I notice another bug that when we normalize a `_StridedShard`'s placement, it will also trigger parent class `Shard`'s `split_tensor` method because it will create a Shard class [here](0c14f55de6/torch/distributed/tensor/_api.py (L783)). I think we never test `distribute_tensor` for `_StridedShard` before. So I added a test here to compare against ordered shard.
Using classmethod because the _split_tensor logic is different between `Shard` and `_StridedShard`. Basically I want to shard on local tensors without initializing the Shard object:
```
local_tensor = _StridedShard._make_shard_tensor(dim, tensor, mesh, mesh_dim, split_factor=split_factor)
local_tensor = Shard._make_shard_tensor(dim, tensor, mesh, mesh_dim)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165533
Approved by: https://github.com/XilunWu
Currently the docs for `torch.mode` include a note:
`This function is not defined for torch.cuda.Tensor yet.`
However with `torch==2.7.1+cu126` when I try to get the mode of a Tensor that is in cuda memory, I do not face any issues:
```
>>> a = torch.tensor([0, 2, 1, 1, 1, 3, 3])
>>> a.mode()
torch.return_types.mode(
values=tensor(1),
indices=tensor(4))
>>> a.cuda().mode()
torch.return_types.mode(
values=tensor(1, device='cuda:0'),
indices=tensor(4, device='cuda:0'))
```
Am I misunderstanding the note? If not, I suggest removing it.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165614
Approved by: https://github.com/mikaylagawarecki
Summary: This starts writing the compiler_config metadata into logger
Test Plan:
Modified existing test case to make sure this is not null.
(Also eyeballed what we're logging tomake sure it's reasonable
Reviewed By: masnesral
Differential Revision: D84014636
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165581
Approved by: https://github.com/masnesral
By adding a few small helpers (e.g., a `splice` method to `_MeshLayout`, and making `_init_process_groups` static and thus stateless) we can substantially shorten the definition of the unflatten method, and help readability.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165556
Approved by: https://github.com/fduwjj
ghstack dependencies: #165554, #165555
The refactoring of DeviceMesh is heavily constrained by the signature of its constructor, which is a public API which contains some "legacy" concepts which we'd love to get rid of, such as an explicit/materialized `mesh` Tensor.
In other languages the solution to this would be to add a private overload of the constructor. Python doesn't natively allow this, but in this PR I managed to build something that approximates it.
This new private constructor basically only takes `_layout`, `_global_rank_permutation`, and `mesh_dim_names`.
With such a constructor we can effectively simplify a lot of callsites and get rid of the `_create_mesh_from_ranks` helper method. That's a good thing because it was instantiating many DeviceMeshes in a for loop, which always felt unnecessary.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165555
Approved by: https://github.com/fduwjj, https://github.com/fegin
ghstack dependencies: #165554
The goal of this PR is to avoid storing the explicit `mesh` Tensor inside each DeviceMesh, and instead compute it on-the-fly when the end user needs it, and try to replace all of its internal usages with `_layout` and the newly-introduced `_global_rank_permutation` Tensor. The name of this attribute is up for debate. The advantage of the `_global_rank_permutation` Tensor is that it is _the same_ Tensor for the root mesh and all its children, so it doesn't need to be copied/reallocated.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165554
Approved by: https://github.com/fduwjj
Fixes#165447
On AOTAutogradCache load, the serialization function we pick is just lambda: self, because the object itself is an AOTAutogradCacheEntry. However, this isn't safe, because `wrap_post_compile` will make `self` unserializable, since it needs to load triton kernels and stuff!
So instead, on AOTAutogradCache load, we preserve the bytes that were used to load the object to begin with, and return that object on a call to serialize(). This effectively makes it so that we save a copy of the pre-hydrated artifact, without needing to do an eager copy until someone actually calls `serialize`.
Test Plan:
Run
```py
import torch
class M(torch.nn.Module):
def __init__(self):
super().__init__()
self.linear1 = torch.nn.Linear(2, 4)
self.relu = torch.nn.ReLU()
self.linear2 = torch.nn.Linear(4, 8)
def forward(self, x):
return self.linear2(self.relu(self.linear1(x)))
device = "cuda"
m = M().to(device)
sample_inputs = (torch.randn(2, 2, device=device),)
eager_out = m(*sample_inputs)
with torch._dynamo.config.patch("enable_aot_compile", True):
compiled_fn_path = "./m.pt"
compiled_fn = torch.compile(
m,
fullgraph=True
).forward.aot_compile((sample_inputs, {}))
compiled_fn.save_compiled_function(compiled_fn_path)
torch._dynamo.reset()
with torch.compiler.set_stance("fail_on_recompile"):
with open(compiled_fn_path, "rb") as f:
loaded_fn = torch.compiler.load_compiled_function(f)
assert loaded_fn is not None
compiled_out = loaded_fn(m, *sample_inputs)
assert torch.allclose(eager_out, compiled_out)
```
twice, see that it succeeds.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165474
Approved by: https://github.com/yiming0416, https://github.com/zhxchen17
