[inductor] Fix lowerings that create unexpected aliases #105173
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/105173
Note: Links to docs will display an error until the docs builds have been completed. ✅ 1 Unrelated FailureAs of commit 45f8c32: UNSTABLE - The following job failed but was likely due to flakiness present on trunk and has been marked as unstable:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This may give the wrong result if the result in some cases, e.g.
```python
torch.compile()
def fn(x):
tmp = x.ceil()
x.add_(10)
return tmp
a = torch.zeros((), dtype=torch.int64)
fn(a) # tensor(10)
```
ghstack-source-id: 1b4683fe003cd65529760df3d9df2b35ed82d5fe
Pull Request resolved: #105173
This may give the wrong result if the result in some cases, e.g.
```python
@torch.compile()
def fn(x):
tmp = x.ceil()
x.add_(10)
return tmp
a = torch.zeros((), dtype=torch.int64)
fn(a) # tensor(10)
```
[ghstack-poisoned]
This may give the wrong result in some cases, e.g.
```python
torch.compile()
def fn(x):
tmp = x.ceil()
x.add_(10)
return tmp
a = torch.zeros((), dtype=torch.int64)
fn(a) # tensor(10)
```
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx ipiszy ngimel yf225 chenyang78 kadeng muchulee8
[ghstack-poisoned]
This may give the wrong result if the result in some cases, e.g.
```python
torch.compile()
def fn(x):
tmp = x.ceil()
x.add_(10)
return tmp
a = torch.zeros((), dtype=torch.int64)
fn(a) # tensor(10)
```
ghstack-source-id: 18f37b54e605d7cdfa109d4b9b6751123fc60212
Pull Request resolved: #105173
This may give the wrong result in some cases, e.g.
```python
torch.compile()
def fn(x):
tmp = x.ceil()
x.add_(10)
return tmp
a = torch.zeros((), dtype=torch.int64)
fn(a) # tensor(10)
```
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx ipiszy ngimel yf225 chenyang78 kadeng muchulee8
[ghstack-poisoned]
This may give the wrong result if the result in some cases, e.g.
```python
torch.compile()
def fn(x):
tmp = x.ceil()
x.add_(10)
return tmp
a = torch.zeros((), dtype=torch.int64)
fn(a) # tensor(10)
```
ghstack-source-id: 110e8d89ae4e49ba75f6513913e4ce640cdcfdbf
Pull Request resolved: #105173
This may give the wrong result if the result in some cases, e.g.
```python
torch.compile()
def fn(x):
tmp = x.ceil()
x.add_(10)
return tmp
a = torch.zeros((), dtype=torch.int64)
fn(a) # tensor(10)
```
ghstack-source-id: 110e8d89ae4e49ba75f6513913e4ce640cdcfdbf
Pull Request resolved: pytorch#105173
This may give the wrong result in some cases, e.g.
```python
torch.compile()
def fn(x):
tmp = x.ceil()
x.add_(10)
return tmp
a = torch.zeros((), dtype=torch.int64)
fn(a) # tensor(10)
```
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx ipiszy ngimel yf225 chenyang78 kadeng muchulee8
[ghstack-poisoned]
This may give the wrong result if the result in some cases, e.g.
```python
torch.compile()
def fn(x):
tmp = x.ceil()
x.add_(10)
return tmp
a = torch.zeros((), dtype=torch.int64)
fn(a) # tensor(10)
```
ghstack-source-id: 7e1aa4a2ac4561711c4464e63a50ea13df475d62
Pull Request resolved: pytorch#105173
| if arg.get_dtype() != dtype: | ||
| return to_dtype(arg, dtype) | ||
| return arg |
There was a problem hiding this comment.
This is the only functional change here. It fixes an infinite recursion issue created by to_dtype calling clone, which is wrapped with type promotion that itself calls to_dtype. This only happens for no-op dtype conversions, where it was wasteful to clone anyway so better to not call to_dtype when the dtypes already match here.
|
@pytorchbot merge |
Merge failedReason: This PR needs a If not, please add the To add a label, you can comment to pytorchbot, for example For more information, see Details for Dev Infra teamRaised by workflow job |
|
@pytorchbot merge |
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
|
Hi @peterbell10, I tried with your PR but it seems the unit test provided in your PR description still gives incorrect result:
In addition, the change in this PR has brought performance regression to several models: As we're approaching the PyTorch 2.1 branch cut, may I know if you could help take a look at this issue? |
|
I forgot to ask you to add the script in the OP as a test. Could you do so while investigating those issues in afollow up @peterbell10 ? |
Stack from ghstack (oldest at bottom):
This may give the wrong result in some cases, e.g.
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @Xia-Weiwen @wenzhe-nrv @jiayisunx @ipiszy @ngimel @yf225 @chenyang78 @kadeng @muchulee8 @aakhundov