[ExecuTorch][Llama][ez] Remove unwrap_tensor_subclass from 4w quantization path#18975
Merged
[ExecuTorch][Llama][ez] Remove unwrap_tensor_subclass from 4w quantization path#18975
Conversation
…ation path D100066455 changed the Llama export pipeline to quantize weights in the checkpoint dtype (typically bfloat16) before casting to the computation dtype (fp32). This introduced a regression for Vulkan 4w export: `dequantize_affine` ops produced bfloat16 outputs, which Vulkan doesn't support, causing the graph to be split into multiple partitions. When `sym_constrain_range_for_size` constraint nodes were partitioned into a different delegate than the `_local_scalar_dense` + `slice_copy` ops they constrain, ExportPass re-tracing (in ConvertToLinearPass, SpecPropPass, etc.) would crash with `GuardOnDataDependentSymNode: Could not guard on data-dependent expression u539 < 0`. The root cause is `unwrap_tensor_subclass()`. This function decomposes `IntxUnpackedToInt8Tensor` into plain tensors via `torch.nn.utils.parametrize`, capturing the subclass's metadata — including its `dtype` attribute (which controls `dequantize_affine`'s output dtype) — as a frozen snapshot in `UnwrapTensorSubclass.rebuild_stack`. A subsequent `model.to(dtype=fp32)` casts the plain tensors but cannot update the frozen metadata, so `dequantize_affine` continues to output bfloat16. `unwrap_tensor_subclass()` was originally needed as a workaround because `torch.export` did not support tensor subclasses natively. This is no longer the case — the 8da4w path already works without it (using the same `IntxUnpackedToInt8Tensor` subclass), and `torch.export` traces through the subclass correctly. Removing it makes 4w consistent with 8da4w and avoids the metadata-freezing issue entirely. This change was authored with Claude. Differential Revision: [D101252037](https://our.internmc.facebook.com/intern/diff/D101252037/) ghstack-source-id: 368594388 Pull Request resolved: #18957
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18975
Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
This PR needs a
|
SS-JIA
approved these changes
Apr 17, 2026
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
2 participants
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.
This PR was created by the merge bot to help merge the original PR into the main branch.
ghstack PR number: #18957 by @SS-JIA
^ Please use this as the source of truth for the PR details, comments, and reviews
ghstack PR base: https://github.com/pytorch/executorch/tree/gh/SS-JIA/521/base
ghstack PR head: https://github.com/pytorch/executorch/tree/gh/SS-JIA/521/head
Merge bot PR base: https://github.com/pytorch/executorch/tree/main
Merge bot PR head: https://github.com/pytorch/executorch/tree/gh/SS-JIA/521/orig
Differential Revision: D101252037
@diff-train-skip-merge