Repository navigation
Keep cat_out's nnlib input arrays off variable-length stack arrays (#24319) - #24319
Open
zonglinpeng wants to merge 2 commits into
Open
zonglinpeng wants to merge 2 commits into
zonglinpeng wants to merge 2 commits into
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24319
Note: Links to docs will display an error until the docs builds have been completed. ❌ 2 New Failures, 2 PendingAs of commit bec367b with merge base 822cb60 ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Contributor
|
@zonglinpeng has exported this pull request. If you are a Meta employee, you can view the originating Diff in D128236590. |
This PR needs a
|
Summary: `rlx_arbitrary_image_stylization` fails in the ExecuTorch emitter: `InternalError: Tensor has a memory_format that is unsupported in ExecuTorch: torch.channels_last` (`exir/tensor.py` `memory_format_enum` raises `KeyError: torch.channels_last`). Its export carries `clone(memory_format=torch.channels_last)` on ops that stay on the CPU. Cadence lowers with `_skip_dim_order=True`, and so does Helios, which runs Cadence's remove passes. Every edge tensor is therefore contiguous and a memory format has no runtime representation. The emitter accepts only contiguous and preserve. The upstream `RemoveCloneOpsTransform` deliberately keeps clones that change memory format, because with dim order they carry a real layout, and Cadence's wrapper `RemoveCloneOpsTransformImported` ran it unchanged. Nothing removed the clone. The wrapper becomes one pass, renamed `RemoveCloneOpsPass`: a `RemoveCloneOpsTransform` subclass that also treats an `aten.clone` with any memory format as a plain copy, so the clone is removed. `_clone_dim_order` keeps the upstream behavior. Dead-code elimination still runs afterwards, as it did in the wrapper. The upstream transform is unchanged. The rename touches every caller in fbcode and xplat: Cadence passes, Helios sharding transforms, Jarvis, the SAS compiler and the xr_sonic gleam export. This replaces the Helios-only `StripCloneMemoryFormat` pass from the earlier version of this diff. Differential Revision: D125145150
…ytorch#24319) Summary: `impl::G3::native::cat_out` declared variable-length arrays sized by the number of inputs (about 76 B per input) before it checked whether the `xa_nn_cat` path applies. A 256-input int64 cat in RT-DETR (`rtdetr_r18vd`, int8) therefore grew the frame by about 19 KB on the DSP thread that runs the operator, overflowed that thread's 7488 B stack and tripped the Zephyr stack sentinel on the first inference. Raising the stack did not help: 8 KB, 11.5 KB and 16 KB stacks all overflowed. The nnlib path now runs only for same-dtype, supported-dtype cats with at most 16 inputs, and uses fixed-size arrays. Larger cats take the existing generic loop. The frame is a fixed 1440 B: the number of dynamic stack adjustments in `cat_out` goes from 11 to 0 in the built DSP image. Adds a 40-input `cat` test case so the fallback branch is covered. The 256 inputs are one tensor repeated: every entry of that cat references the same value and buffer. RT-DETR's decoder gathers with `topk_ind.unsqueeze(-1).repeat(1, 1, d_model)`, and Cadence's `ReplaceRepeatWithCatPass` lowers a `repeat` as a cat of the same tensor, so `d_model` = 256 becomes a 256-entry cat. The kernel is still the place to fix the overflow: the arrays are sized by entry count, not by distinct tensors, and a 100-entry cat of distinct inputs (EfficientDet-Lite0's full graph) would need about 7.6 KB of them, more than the whole 7488 B stack. Lowering `repeat` as a broadcast copy would remove RT-DETR's long list but leave genuinely wide cats exposed, so it is at most a later optimization. Differential Revision: D128236590
meta-codesync
Bot
changed the title
zonglinpeng
force-pushed
the
export-D128236590
branch
from
November 2, 2026 18:33
ea355db to
bec367b
Compare
This branch was previously deployed
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
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.
Summary:
impl::G3::native::cat_outdeclared variable-length arrays sized by the number of inputs (about 76 B per input) before it checked whether thexa_nn_catpath applies. A 256-input int64 cat in RT-DETR (rtdetr_r18vd, int8) therefore grew the frame by about 19 KB on the DSP thread that runs the operator, overflowed that thread's 7488 B stack and tripped the Zephyr stack sentinel on the first inference. Raising the stack did not help: 8 KB, 11.5 KB and 16 KB stacks all overflowed.The nnlib path now runs only for same-dtype, supported-dtype cats with at most 16 inputs, and uses fixed-size arrays. Larger cats take the existing generic loop. The frame is a fixed 1440 B: the number of dynamic stack adjustments in
cat_outgoes from 11 to 0 in the built DSP image.Adds a 40-input
cattest case so the fallback branch is covered.The 256 inputs are one tensor repeated: every entry of that cat references the same value and buffer. RT-DETR's decoder gathers with
topk_ind.unsqueeze(-1).repeat(1, 1, d_model), and Cadence'sReplaceRepeatWithCatPasslowers arepeatas a cat of the same tensor, sod_model= 256 becomes a 256-entry cat. The kernel is still the place to fix the overflow: the arrays are sized by entry count, not by distinct tensors, and a 100-entry cat of distinct inputs (EfficientDet-Lite0's full graph) would need about 7.6 KB of them, more than the whole 7488 B stack. Loweringrepeatas a broadcast copy would remove RT-DETR's long list but leave genuinely wide cats exposed, so it is at most a later optimization.Differential Revision: D128236590