remove guard_size_oblivious from unbind. #148815
Open
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
[ghstack-poisoned]
This was referenced Mar 8, 2025
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/148815
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ❌ 4 New FailuresAs of commit 8260058 with merge base c4a0b31 ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This PR needs a
|
This was referenced Mar 10, 2025
laithsakka
commented
Mar 10, 2025
This was interesting why?
i initially thought this is the best rewrite for this
```
# Looking at this what will happen because of guard_size_oblivious/guard_or_false
# At least for normal tesnors we would throw during runtime when t.shape[dim] == 0 the following error.
# RuntimeError: number of sections must be larger than 0, got 0
# Because of calling torch.tensor_split(t, t.shape[dim], dim)
# Maybe a better error msg alternative here, is to throw in such case an error
# RuntimeError: unbind unbacked semantics does not support 0 inputs during runtime.
def fall_back():
torch._check(t.shape[dim] == 0,
lambda: "unbind unbacked semantics does not support 0 inputs during runtime.")
return False
if guard_or_else(t.shape[dim] == 0, fall_back):
return ()
else:
return tuple(
torch.squeeze(s, dim) for s in torch.tensor_split(t, t.shape[dim], dim)
)
```
but then i tried to add examples
ex:
```
from torch.fx.experimental.symbolic_shapes import sym_or
torch.compile(fullgraph=True)
def func(y):
return y.unbind(dim=2)
z = torch.ones(1,1,1,1)
func(z)
torch._dynamo.decorators.mark_unbacked(z, 2)
```
but we actually do hit another guard
Specifically guard_int() bellow
```
std::vector<Tensor> tensor_split_sections_symint(
const Tensor& self,
c10::SymInt sym_sections,
int64_t dim) {
TORCH_CHECK(
self.dim() > 0,
"tensor_split expected at least a 1-dimensional tensor, but got a tensor with ",
self.dim(),
" dims");
int64_t dim_ = maybe_wrap_dim(dim, self.dim());
// NB: intentional, sections specifies number of output tensors, which
// cannot be polymorphic
int64_t sections = sym_sections.guard_int(__FILE__, __LINE__);
```
so my thoughts, is i do not know what is being avoid by adding this, it was added in its on PR so assume for a reason?
#124959
I will keep the current semantics of guard_size_oblivious and just use guard_or_false.
maybe for other types of tensors tensor_split_sections_symint does not throw runtime error or re-guard?
idk
[ghstack-poisoned]
Divigroup-RAP
pushed a commit
to Divigroup-RAP/PYTORCH
that referenced
this pull request
Apr 22, 2025
ghstack-source-id: 3dd8fef Pull Request resolved: pytorch/pytorch#148815
This was interesting why?
i initially thought this is the best rewrite for this
```
# Looking at this what will happen because of guard_size_oblivious/guard_or_false
# At least for normal tesnors we would throw during runtime when t.shape[dim] == 0 the following error.
# RuntimeError: number of sections must be larger than 0, got 0
# Because of calling torch.tensor_split(t, t.shape[dim], dim)
# Maybe a better error msg alternative here, is to throw in such case an error
# RuntimeError: unbind unbacked semantics does not support 0 inputs during runtime.
def fall_back():
torch._check(t.shape[dim] == 0,
lambda: "unbind unbacked semantics does not support 0 inputs during runtime.")
return False
if guard_or_else(t.shape[dim] == 0, fall_back):
return ()
else:
return tuple(
torch.squeeze(s, dim) for s in torch.tensor_split(t, t.shape[dim], dim)
)
```
but then i tried to add examples
ex:
```
from torch.fx.experimental.symbolic_shapes import sym_or
torch.compile(fullgraph=True)
def func(y):
return y.unbind(dim=2)
z = torch.ones(1,1,1,1)
func(z)
torch._dynamo.decorators.mark_unbacked(z, 2)
```
but we actually do hit another guard
Specifically guard_int() bellow
```
std::vector<Tensor> tensor_split_sections_symint(
const Tensor& self,
c10::SymInt sym_sections,
int64_t dim) {
TORCH_CHECK(
self.dim() > 0,
"tensor_split expected at least a 1-dimensional tensor, but got a tensor with ",
self.dim(),
" dims");
int64_t dim_ = maybe_wrap_dim(dim, self.dim());
// NB: intentional, sections specifies number of output tensors, which
// cannot be polymorphic
int64_t sections = sym_sections.guard_int(__FILE__, __LINE__);
```
so my thoughts, is i do not know what is being avoid by adding this, it was added in its on PR so assume for a reason?
#124959
I will keep the current semantics of guard_size_oblivious and just use guard_or_false.
maybe for other types of tensors tensor_split_sections_symint does not throw runtime error or re-guard?
idk
[ghstack-poisoned]
pianpwk
reviewed
May 2, 2025
|
|
||
| # Note: t.shape[dim] can't be dynamic or unbacked, even if we use guard_or_false here we will fail | ||
| # later in the split since t.shape[dim] control the number of output tensors. | ||
| if t.shape[dim] == 0: |
There was a problem hiding this comment.
I haven't run this locally, but it seems the data-dependent error the user would see would now originate from here? Their first instinct might then be to try to slap a guard_size_oblivious on again.
Is there some way we can prevent that? Not that it's a big deal.
e.g. by adding
n_out = int(t.shape[dim]) # unbind returns a list, so we force-specialize on the output length anyways; this cannot remain symbolic
...
return tuple(
torch.squeeze(s, dim) for s in torch.tensor_split(t, n_out, dim)
)
There was a problem hiding this comment.
yeh I added this comment so people avoid that, good idea i can do that to make it explicit.
pianpwk
approved these changes
May 2, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Stack from ghstack (oldest at bottom):
unbind will always specialize on dim, because it determine the number of output tensors.
guard_size_oblivious is not useful there and more confusing probably for code readers
added a comment and a test that verifies the specialization.