Evgeni Burovski
324b23f337
MAINT: torch/_numpy: remove stubs raising NIError ( #108902 )
...
Remove remaining stubs. There is no point raising NotImplementedError, now that a missing function triggers a graph break just by being missing in `torch._numpy` namespace.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108902
Approved by: https://github.com/lezcano
2023-09-09 00:11:14 +00:00
Evgeni Burovski
1f20531939
fall back to eager on NotImplementedError ( #107863 )
...
Follow-up to https://github.com/pytorch/pytorch/pull/107710 :
Help dynamo fall back to eager when compiling unimplemented numpy constructs:
- arrays of strings
- (arg){min, max} for complex types
- various arguments typed as NotImplemented (`np.ones(4, order="F")` etc)
- numpy functions which torch._numpy does not implement
To test, run (we do not implement arrays of strings)
```
import torch
import numpy as np
@torch.compile(fullgraph=False)
def fn():
return np.asarray(["L", "U"])
```
and observe it compiles with fullgraph=False and fails with fullgraph=True
Fixes https://github.com/pytorch/pytorch/issues/107970
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107863
Approved by: https://github.com/ezyang , https://github.com/lezcano
2023-09-07 21:22:20 +00:00
lezcano
2a6ef9b04d
[dynamo] Avoid recompilation when the PyTorch function accepts scalars ( #108162 )
...
Before, it would create a 0D tensor with the input, which would incur in
a guard and specialisation.
It's not clear whether the guard and specialisation is the right behaviour
when we create 0D tensors, but that's a story for another day.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108162
Approved by: https://github.com/ev-br , https://github.com/peterbell10
2023-09-01 14:35:42 +00:00
Jirka Borovec
9178deedff
removing some redundant str splits ( #106089 )
...
drop some redundant string splits, no factual changes, just cleaning the codebase
Pull Request resolved: https://github.com/pytorch/pytorch/pull/106089
Approved by: https://github.com/albanD , https://github.com/malfet
2023-09-01 00:22:58 +00:00
Evgeni Burovski
01dfa7620d
MAINT: np.unique works with f16 directly ( #108228 )
...
(follow up on gh-107768)
Remove a f16->f32 workaround from np.unique, since torch.unique and np.unique seem to just work with float16 tensors.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108228
Approved by: https://github.com/lezcano
2023-08-31 16:21:13 +00:00
Evgeni Burovski
55d6b80188
torch._numpy: keep f16 CUDA tensors in f16 where possible ( #107768 )
...
Contain workarounds for _RuntimeError: "addmm_impl_cpu_" not implemented for 'Half'_ to CPU tensors, do computations on CUDA tensors in f16.
Fixes https://github.com/Quansight-Labs/numpy_pytorch_interop/issues/170
We do not really systematically test CUDA tensors in torch._numpy, so I only spot-checked locally that the affected functions work with `tensor.to("cuda")`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107768
Approved by: https://github.com/lezcano
2023-08-23 18:35:47 +00:00
lezcano
fada0527fa
Dispatch take_along_axis to gather ( #107711 )
...
Gather does the same thing, but it's much better supported in the
`torch.compile` stack
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107711
Approved by: https://github.com/ezyang
ghstack dependencies: #107687 , #107688 , #107710
2023-08-23 01:21:23 +00:00
lezcano
62113a2361
[dynamo] np.sort(complex) is not implemented ( #107710 )
...
This issue was discovered once we were able to trace without breaking
in https://github.com/pytorch/pytorch/pull/107689 . Same for the next
one.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107710
Approved by: https://github.com/ezyang
ghstack dependencies: #107687 , #107688
2023-08-23 01:21:23 +00:00
Evgeni Burovski
da67b414d9
torch._numpy: remove noops and half-implemented nan-functions ( #107596 )
...
As discussed in the review of https://github.com/pytorch/pytorch/pull/106211 , remove several noops (https://github.com/pytorch/pytorch/pull/106211#pullrequestreview-1559806543 and https://github.com/pytorch/pytorch/pull/106211#pullrequestreview-1559809287 ).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107596
Approved by: https://github.com/lezcano
2023-08-21 21:17:55 +00:00
lezcano
a9dca53438
NumPy support in torch.compile ( #106211 )
...
RFC: https://github.com/pytorch/rfcs/pull/54
First commit is the contents of https://github.com/Quansight-Labs/numpy_pytorch_interop/
We have already been using this in core for the last few months as a external dependency. This PR pulls all these into core.
In the next commits, I do a number of things in this order
- Fix a few small issues
- Make the tests that this PR adds pass
- Bend backwards until lintrunner passes
- Remove the optional dependency on `torch_np` and simply rely on the upstreamed code
- Fix a number dynamo tests that were passing before (they were not tasting anything I think) and are not passing now.
Missing from this PR (but not blocking):
- Have a flag that deactivates tracing NumPy functions and simply breaks. There used to be one but after the merge stopped working and I removed it. @lezcano to investigate.
- https://github.com/pytorch/pytorch/pull/106431#issuecomment-1667079543 . @voznesenskym to submit a fix after we merge.
All the tests in `tests/torch_np` take about 75s to run.
This was a work by @ev-br, @rgommers @honno and I. I did not create this PR via ghstack (which would have been convenient) as this is a collaboration, and ghstack doesn't allow for shared contributions.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/106211
Approved by: https://github.com/ezyang
2023-08-11 00:39:32 +00:00
