[dynamic shapes] aten.constant_pad_nd meta impl #152129
Closed
+63
−0
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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/152129
Note: Links to docs will display an error until the docs builds have been completed. ✅ You can merge normally! (1 Unrelated Failure)As of commit 304fddd with merge base 89c0c3c ( BROKEN TRUNK - The following job failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
| @@ -4269,6 +4269,26 @@ def forward(self, x): | |||
| ): | |||
| _ = export(M(), (torch.tensor([2, 3, 5]),)) | |||
|
|
|||
| @testing.expectedFailureTrainingIRToRunDecomp | |||
8000
There was a problem hiding this comment.
Whats the failure? Since this is the default path, i think we should fix whatever it shows.
|
in the summary you mention "we know this always produces a clone" |
laithsakka
reviewed
Apr 29, 2025
| @@ -7322,6 +7323,29 @@ def softmax(x: Tensor, dim: int, half_to_float: bool) -> Tensor: | |||
| return res | |||
|
|
|||
|
|
|||
| @register_meta(aten.constant_pad_nd) | |||
| @out_wrapper() | |||
| def _constant_pad_nd_meta(input, pad, value=0): | |||
There was a problem hiding this comment.
shall we add the torch checks from the decomp version in order to fail earlier?
laithsakka
reviewed
Apr 29, 2025
There was a problem hiding this comment.
I reviewed this by comparing to
# This replicates at::constant_pad_nd, defined in ATen/native/PadNd.cpp
@register_decomposition(aten.constant_pad_nd)
@out_wrapper()
def constant_pad_nd(
input: TensorLikeType, pad: list[int], value: NumberType = 0
) -> TensorLikeType:
torch._check(
len(pad) % 2 == 0,
lambda: f"Length of pad must be even but instead it equals {len(pad)}",
)
input_sizes = input.shape
l_inp = len(input_sizes)
l_pad = len(pad) // 2
l_diff = l_inp - l_pad
torch._check(
l_inp >= l_pad,
lambda: "Length of pad should be no more than twice the number of "
f"dimensions of the input. Pad length is {len(pad)} while the input has "
f"{l_inp} dimensions.",
)
c_input = input
for i in range(l_diff, l_inp):
pad_idx = 2 * (l_inp - i - 1)
if pad[pad_idx] < 0:
c_input = c_input.narrow(i, -pad[pad_idx], c_input.shape[i] + pad[pad_idx])
if pad[pad_idx + 1] < 0:
c_input = c_input.narrow(i, 0, c_input.shape[i] + pad[pad_idx + 1])
# If all the pads are negative we can return the result.
# Avoid early exiting if all pads = 0 to prevent specialization on export.
# During export, raw if statements are specialized on the input, meaning
# that we lose a branch depending on the example input used to export.
# Here, this is either the case where all pads = 0, or the case where at
# least one pad > 0 and the rest are >= 0.
# Avoiding the early exit when all pads = 0 ensures we can export
# constant_pad_nd for cases when all pads >= 0.
# Note: if any pads are negative, this code specializes due to the if statements above.
if builtins.all(p < 0 for p in pad):
return c_input.clone()
new_shape = list(input_sizes[:l_diff])
for i in range(l_pad):
pad_idx = len(pad) - ((i + 1) * 2)
new_dim = input_sizes[l_diff + i] + pad[pad_idx] + pad[pad_idx + 1]
torch._check(
new_dim > 0,
lambda: f"The input size {input_sizes[l_diff + i]}, plus negative padding "
f"{pad[pad_idx]} and {pad[pad_idx + 1]} resulted in a negative output size, "
f"which is invalid. Check dimension {l_diff + i} of your input.",
)
new_shape.append(new_dim)
memory_format = utils.suggest_memory_format(input)
output = torch.empty(
new_shape,
dtype=input.dtype,
device=input.device,
requires_grad=input.requires_grad,
memory_format=memory_format,
)
if value == 0 and input.dtype == torch.bool:
value = False
# torch.fill isn't typed to allow complex values
output = torch.fill(output, value) # type: ignore[arg-type]
c_output = output
for i in range(l_diff, l_inp):
pad_idx = 2 * (l_inp - i - 1)
if pad[pad_idx] >= 0:
c_output = c_output.narrow(
i, pad[pad_idx], c_output.shape[i] - pad[pad_idx]
)
if pad[pad_idx + 1] >= 0:
c_output = c_output.narrow(i, 0, c_output.shape[i] - pad[pad_idx + 1])
prims.copy_to(c_output, c_input)
return output
looks legit can we add the torch checks though
laithsakka
approved these changes
Apr 29, 2025
just that even if all padding is zero or negative, we'll never alias the original tensor, so this meta-kernel isn't semantic changing |
|
@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 |
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.
We know the output shape, and we know this always produces a clone. Avoids data-dependent errors from the decomposition.
along with #150483, should fix #123855