Repository navigation
Conversation
meta-cla
Bot
added
the
CLA Signed
Contributor
|
@pssrawat has exported this pull request. If you are a Meta employee, you can view the originating Diff in D128001279. |
meta-codesync
Bot
changed the title
pssrawat
force-pushed
the
export-D128001279
branch
from
October 29, 2026 06:26
b510da9 to
0b139c1
Compare
pssrawat
added a commit
to pssrawat/executorch
that referenced
this pull request
Oct 29, 2026
Summary: On the explicit-layout pass list, a Conv1d is lowered as a Conv2d, so each convolution's NHWC output goes through a rank-4 layout copy followed by a squeeze back to rank 3. `RemovePermutesAroundElementwiseOps` cancels permutes around a region of elementwise ops, but it can only cross a squeeze or unsqueeze while walking downstream. When a region walks upstream into a neighbouring convolution's output, for example through the second operand of a residual add, it reaches that squeeze from below, carrying a rank-3 permutation with a rank-4 permute behind it. It discards the whole region, and the layout copies between consecutive convolutions survive as transposes. Add `CortexMRemovePermutesAroundElementwiseOps`, which sinks each layout copy below a following squeeze and then runs `RemovePermutesAroundElementwiseOps`. The sinking is done by `CortexMPostponeLayoutCopyBelowSqueezePass`, a restricted `PostponePermuteOpBelowSqueezeOrUnsqueezeLikeView`: it only moves the layout copies this pipeline inserts, leaving user permutes in place, and never moves one below an unsqueeze, which would raise its rank, up to a rank-5 layout copy that Cortex-M cannot run. After sinking, a convolution output is a squeeze followed by a rank-3 layout copy, so regions meet copies of their own rank and cancel as they do for Conv2d. The sinking runs inside the pass, outside the pass manager's dtype-preserving retrace, so when only the sinking changed the graph the pass retraces once more to restore the folded int8 dtypes. Authored with an AI coding assistant (Claude). Differential Revision: D128001279
pssrawat
added a commit
to pssrawat/executorch
that referenced
this pull request
Oct 29, 2026
Summary: On the explicit-layout pass list, a Conv1d is lowered as a Conv2d, so each convolution's NHWC output goes through a rank-4 layout copy followed by a squeeze back to rank 3. `RemovePermutesAroundElementwiseOps` cancels permutes around a region of elementwise ops, but it can only cross a squeeze or unsqueeze while walking downstream. When a region walks upstream into a neighbouring convolution's output, for example through the second operand of a residual add, it reaches that squeeze from below, carrying a rank-3 permutation with a rank-4 permute behind it. It discards the whole region, and the layout copies between consecutive convolutions survive as transposes. Add `CortexMRemovePermutesAroundElementwiseOps`, which sinks each layout copy below a following squeeze and then runs `RemovePermutesAroundElementwiseOps`. The sinking is done by `CortexMPostponeLayoutCopyBelowSqueezePass`, a restricted `PostponePermuteOpBelowSqueezeOrUnsqueezeLikeView`: it only moves the layout copies this pipeline inserts, leaving user permutes in place, and never moves one below an unsqueeze, which would raise its rank, up to a rank-5 layout copy that Cortex-M cannot run. After sinking, a convolution output is a squeeze followed by a rank-3 layout copy, so regions meet copies of their own rank and cancel as they do for Conv2d. The sinking runs inside the pass, outside the pass manager's dtype-preserving retrace, so when only the sinking changed the graph the pass retraces once more to restore the folded int8 dtypes. Authored with an AI coding assistant (Claude). Differential Revision: D128001279
pssrawat
force-pushed
the
export-D128001279
branch
2 times, most recently
from
October 30, 2026 06:11
fe17fb8 to
3d9b0f8
Compare
pssrawat
added a commit
to pssrawat/executorch
that referenced
this pull request
Oct 30, 2026
Summary: On the explicit-layout pass list, a Conv1d is lowered as a Conv2d, so each convolution's NHWC output goes through a rank-4 layout copy followed by a squeeze back to rank 3. `RemovePermutesAroundElementwiseOps` cancels permutes around a region of elementwise ops, but it can only cross a squeeze or unsqueeze while walking downstream. When a region walks upstream into a neighbouring convolution's output, for example through the second operand of a residual add, it reaches that squeeze from below, carrying a rank-3 permutation with a rank-4 permute behind it. It discards the whole region, and the layout copies between consecutive convolutions survive as transposes. Add `CortexMPostponeLayoutCopyBelowSqueezePass`, a restricted `PostponePermuteOpBelowSqueezeOrUnsqueezeLikeView`, and run it just before `RemovePermutesAroundElementwiseOps` in the explicit-layout pass list. It only moves the layout copies this pipeline inserts (its `targets` is `LAYOUT_PERMUTE_COPY`), leaving user permutes in place, and never moves one below an unsqueeze, which would raise its rank, up to a rank-5 layout copy that Cortex-M cannot run. After sinking, a convolution output is a squeeze followed by a rank-3 layout copy, so regions meet copies of their own rank and cancel as they do for Conv2d. As its own pass-list entry, the pass gets the pass manager's dtype-preserving retrace, so the folded int8 dtypes survive the sinking. Authored with an AI coding assistant (Claude). Differential Revision: D128001279
Summary: Q/DQ folding (`FoldAndAnnotateQParamsPass`) always folds the quantize/dequantize pair around a constant-producing op (`full`, `full_like`, `arange`, `eye`, `linspace`, `scalar_tensor`) into the op: it removes the Q node and records the output qparams, but the op itself keeps producing float. The Arm backend follows folding with `ComputeConstantOpsAOTPass`, which evaluates such ops in float and quantizes the result. Cortex-M did not, so in `x + torch.full(shape, 0.5)` the int8 `quantized_add` kernel read a float32 constant, and the output was silently wrong. Add `ComputeQuantizedConstantOpsPass` and run it after folding on both Cortex-M pass lists. - a. It only touches nodes with output qparams, so unquantized constant ops are unchanged. A quantized `full` stays a runtime fill, with its fill value quantized and its dtype set to the qparams' dtype, so a large fill does not turn into a large constant; - b.`full_like` is first converted to `full` with the Arm backend's `ConvertFullLikeToFullPass`. This also fixes a test that was already failing. `surrounded_quantized_op_constant` in `misc/test_quantization.py` was marked as an expected failure. Its `torch.ones(2, 2)` constant is lowered to a `full` op, so the int8 add read float data, and every output element was off by about 0.49. With this change the case matches the quantized reference exactly, so its `xfail` is removed. The other constants in that file are stored tensors, which folding already quantized correctly. Differential Revision: D128002647
…pytorch#24153) Summary: `SharedQspecQuantizer` gives each shared-qparams cluster one set of qparams for every float input and output of its ops. `view_as(x, other)` is in that list, and the cluster was built from all of its inputs, so `other` was pulled in even though it contributes only its shape. `x` and `other` were then quantized with one scale, set by whichever has the wider range: the data loses precision, or the shape operand is clipped. For example, in `(x * y).view_as(other) + other` with `other = z + z`, `x * y` spans about ±2.5 and `other` about ±51. `x * y` was quantized at `other`'s scale of 0.349, about 18x coarser than the 0.0195 it gets when the same reshape is written `reshape(4, 8)`, and the output error against float rose from 0.258 to 0.307. `expand_as` has the same shape-only second operand but was not in the shared-qparams list at all, so every `expand_as` broadcast stayed in float behind a dequantize and a requantize. Shape-as ops (`view_as`, `expand_as`) now share qparams only through their data operand, `args[0]`. The cluster walk does not cross the edge to the shape operand in either direction, and that operand gets no shared input qspec, so it keeps the qparams of its own producer. `expand_as` joins the shared-qparams ops, so like `expand` it is quantized to int8 when its neighbours are, and lowers on Cortex-M to an int8 `expand_copy`. Authored with Claude Code. Differential Revision: D127380012
Summary: On the explicit-layout pass list, a Conv1d is lowered as a Conv2d, so each convolution's NHWC output goes through a rank-4 layout copy followed by a squeeze back to rank 3. `RemovePermutesAroundElementwiseOps` cancels permutes around a region of elementwise ops, but it can only cross a squeeze or unsqueeze while walking downstream. When a region walks upstream into a neighbouring convolution's output, for example through the second operand of a residual add, it reaches that squeeze from below, carrying a rank-3 permutation with a rank-4 permute behind it. It discards the whole region, and the layout copies between consecutive convolutions survive as transposes. Add `CortexMPostponeLayoutCopyBelowSqueezePass`, a restricted `PostponePermuteOpBelowSqueezeOrUnsqueezeLikeView`, and run it just before `RemovePermutesAroundElementwiseOps` in the explicit-layout pass list. It only moves the layout copies this pipeline inserts (its `targets` is `LAYOUT_PERMUTE_COPY`), leaving user permutes in place, and never moves one below an unsqueeze, which would raise its rank, up to a rank-5 layout copy that Cortex-M cannot run. After sinking, a convolution output is a squeeze followed by a rank-3 layout copy, so regions meet copies of their own rank and cancel as they do for Conv2d. As its own pass-list entry, the pass gets the pass manager's dtype-preserving retrace, so the folded int8 dtypes survive the sinking. Authored with an AI coding assistant (Claude). Differential Revision: D128001279
pssrawat
force-pushed
the
export-D128001279
branch
from
October 30, 2026 07:26
3d9b0f8 to
9d6d28d
Compare
This branch has not been 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:
On the explicit-layout pass list, a Conv1d is lowered as a Conv2d, so each convolution's NHWC output goes through a rank-4 layout copy followed by a squeeze back to rank 3.
RemovePermutesAroundElementwiseOpscancels permutes around a region of elementwise ops, but it can only cross a squeeze or unsqueeze while walking downstream. When a region walks upstream into a neighbouring convolution's output, for example through the second operand of a residual add, it reaches that squeeze from below, carrying a rank-3 permutation with a rank-4 permute behind it. It discards the whole region, and the layout copies between consecutive convolutions survive as transposes.Add
CortexMPostponeLayoutCopyBelowSqueezePass, a restrictedPostponePermuteOpBelowSqueezeOrUnsqueezeLikeView, and run it just beforeRemovePermutesAroundElementwiseOpsin the explicit-layout pass list. It only moves the layout copies this pipeline inserts (itstargetsisLAYOUT_PERMUTE_COPY), leaving user permutes in place, and never moves one below an unsqueeze, which would raise its rank, up to a rank-5 layout copy that Cortex-M cannot run. After sinking, a convolution output is a squeeze followed by a rank-3 layout copy, so regions meet copies of their own rank and cancel as they do for Conv2d. As its own pass-list entry, the pass gets the pass manager's dtype-preserving retrace, so the folded int8 dtypes survive the sinking.Authored with an AI coding assistant (Claude).
Differential Revision: D128001279