Repository navigation
Support int16 in generic quantized_layer_norm (#24093) - #24093
Merged
meta-codesync[bot] merged 2 commits intoOct 23, 2026
Merged
Conversation
mvartani-meta
requested review from
larryliu0820,
mcremon-meta and
rascani
as code owners
October 22, 2026 20:21
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24093
Note: Links to docs will display an error until the docs builds have been completed. ❌ 2 Unclassified FailuresAs of commit 25db302 with merge base 2c3e231 ( UNCLASSIFIED FAILURES - DrCI could not classify the following jobs because the workflow did not run on the merge base. The failures may be pre-existing on trunk or introduced by this PR:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Contributor
|
@mvartani-meta has exported this pull request. If you are a Meta employee, you can view the originating Diff in D126951724. |
This PR needs a
|
DrJessop
approved these changes
Oct 22, 2026
meta-codesync
Bot
changed the title
mvartani-meta
added a commit
to mvartani-meta/executorch
that referenced
this pull request
Oct 22, 2026
Summary: The generic quantized_layer_norm kernel has the same int16 gap as the HiFi kernel fixed earlier in this stack. The generic kernel is the portable implementation used on CPU and on Xtensa cores that have no backend-specific layer norm kernel. Before this change, both overloads dispatched with ET_FORALL_CADENCE_QUANTIZED_TYPES, so int16 fell through to an ET_DCHECK_MSG that compiles out in release builds and left the output unwritten. The statistics also used int32 accumulators, which overflow for int16 because one squared value can reach 2^30. The variance was formed as E[x^2] - E[x]^2, which cancels catastrophically when the mean is large compared to the spread. This applies the same fix as the HiFi kernel: - Dispatch both overloads with ET_FORALL_CADENCE_QUANTIZED_TYPES_WITH_INT16. - Accumulate in int64 for 16-bit types; 8-bit keeps int32 in the inner loop. - Widen to the accumulator type before squaring, so the product cannot overflow int32. - Apply the zero-point correction in int64 and compute the variance numerator N*sum(v^2) - sum(v)^2 exactly in integers. The per-row mean and variance stay in double as before. - Add a gtest with the same hand-computed cases as the HiFi test. Reviewed By: DrJessop Differential Revision: D126951724
mvartani-meta
force-pushed
the
export-D126951724
branch
from
October 22, 2026 22:00
14e38b9 to
ddc8e6b
Compare
mvartani-meta
added a commit
to mvartani-meta/executorch
that referenced
this pull request
Oct 22, 2026
Summary: The generic quantized_layer_norm kernel has the same int16 gap as the HiFi kernel fixed earlier in this stack. The generic kernel is the portable implementation used on CPU and on Xtensa cores that have no backend-specific layer norm kernel. Before this change, both overloads dispatched with ET_FORALL_CADENCE_QUANTIZED_TYPES, so int16 fell through to an ET_DCHECK_MSG that compiles out in release builds and left the output unwritten. The statistics also used int32 accumulators, which overflow for int16 because one squared value can reach 2^30. The variance was formed as E[x^2] - E[x]^2, which cancels catastrophically when the mean is large compared to the spread. This applies the same fix as the HiFi kernel: - Dispatch both overloads with ET_FORALL_CADENCE_QUANTIZED_TYPES_WITH_INT16. - Accumulate in int64 for 16-bit types; 8-bit keeps int32 in the inner loop. - Widen to the accumulator type before squaring, so the product cannot overflow int32. - Apply the zero-point correction in int64 and compute the variance numerator N*sum(v^2) - sum(v)^2 exactly in integers. The per-row mean and variance stay in double as before. - Add a gtest with the same hand-computed cases as the HiFi test. Reviewed By: DrJessop Differential Revision: D126951724
Summary: The HiFi quantized_layer_norm kernels dispatched with ET_FORALL_CADENCE_QUANTIZED_TYPES, so int16 inputs fell through to ET_DCHECK_MSG. That check compiles out in release builds, which left the output tensor unwritten and produced garbage numerics for int16 models. Simply widening the dispatch is not enough because the statistics used int32 accumulators. For int16, a single squared value can reach 2^30, so sq_sum overflows after two elements. The variance was also computed as E[x^2] - E[x]^2 in fp32, which cancels catastrophically when the mean is large relative to the spread. Changes: - Dispatch both overloads with ET_FORALL_CADENCE_QUANTIZED_TYPES_WITH_INT16. - Accumulate sum and sq_sum in int64 for 16-bit types. 8-bit types keep int32 in the inner loop. - Widen to the accumulator type before squaring, so the product cannot overflow int32. - Apply the zero-point correction in int64 and compute the variance numerator N*sum(v^2) - sum(v)^2 exactly in integers before converting to float. For 8-bit this is equal or more accurate than before. - Declare quantized_layer_norm_out and quantized_layer_norm_per_tensor_out in operators.h and add a gtest. Reviewed By: DrJessop Differential Revision: D126939431
Summary: The generic quantized_layer_norm kernel has the same int16 gap as the HiFi kernel fixed earlier in this stack. The generic kernel is the portable implementation used on CPU and on Xtensa cores that have no backend-specific layer norm kernel. Before this change, both overloads dispatched with ET_FORALL_CADENCE_QUANTIZED_TYPES, so int16 fell through to an ET_DCHECK_MSG that compiles out in release builds and left the output unwritten. The statistics also used int32 accumulators, which overflow for int16 because one squared value can reach 2^30. The variance was formed as E[x^2] - E[x]^2, which cancels catastrophically when the mean is large compared to the spread. This applies the same fix as the HiFi kernel: - Dispatch both overloads with ET_FORALL_CADENCE_QUANTIZED_TYPES_WITH_INT16. - Accumulate in int64 for 16-bit types; 8-bit keeps int32 in the inner loop. - Widen to the accumulator type before squaring, so the product cannot overflow int32. - Apply the zero-point correction in int64 and compute the variance numerator N*sum(v^2) - sum(v)^2 exactly in integers. The per-row mean and variance stay in double as before. - Add a gtest with the same hand-computed cases as the HiFi test. Reviewed By: DrJessop Differential Revision: D126951724
mvartani-meta
force-pushed
the
export-D126951724
branch
from
October 22, 2026 22:24
ddc8e6b to
25db302
Compare
mvartani-meta
added a commit
to mvartani-meta/executorch
that referenced
this pull request
Oct 22, 2026
Summary: The generic quantized_layer_norm kernel has the same int16 gap as the HiFi kernel fixed earlier in this stack. The generic kernel is the portable implementation used on CPU and on Xtensa cores that have no backend-specific layer norm kernel. Before this change, both overloads dispatched with ET_FORALL_CADENCE_QUANTIZED_TYPES, so int16 fell through to an ET_DCHECK_MSG that compiles out in release builds and left the output unwritten. The statistics also used int32 accumulators, which overflow for int16 because one squared value can reach 2^30. The variance was formed as E[x^2] - E[x]^2, which cancels catastrophically when the mean is large compared to the spread. This applies the same fix as the HiFi kernel: - Dispatch both overloads with ET_FORALL_CADENCE_QUANTIZED_TYPES_WITH_INT16. - Accumulate in int64 for 16-bit types; 8-bit keeps int32 in the inner loop. - Widen to the accumulator type before squaring, so the product cannot overflow int32. - Apply the zero-point correction in int64 and compute the variance numerator N*sum(v^2) - sum(v)^2 exactly in integers. The per-row mean and variance stay in double as before. - Add a gtest with the same hand-computed cases as the HiFi test. Reviewed By: DrJessop Differential Revision: D126951724
mvartani-meta
added a commit
to mvartani-meta/executorch
that referenced
this pull request
Oct 22, 2026
Summary: The generic quantized_layer_norm kernel has the same int16 gap as the HiFi kernel fixed earlier in this stack. The generic kernel is the portable implementation used on CPU and on Xtensa cores that have no backend-specific layer norm kernel. Before this change, both overloads dispatched with ET_FORALL_CADENCE_QUANTIZED_TYPES, so int16 fell through to an ET_DCHECK_MSG that compiles out in release builds and left the output unwritten. The statistics also used int32 accumulators, which overflow for int16 because one squared value can reach 2^30. The variance was formed as E[x^2] - E[x]^2, which cancels catastrophically when the mean is large compared to the spread. This applies the same fix as the HiFi kernel: - Dispatch both overloads with ET_FORALL_CADENCE_QUANTIZED_TYPES_WITH_INT16. - Accumulate in int64 for 16-bit types; 8-bit keeps int32 in the inner loop. - Widen to the accumulator type before squaring, so the product cannot overflow int32. - Apply the zero-point correction in int64 and compute the variance numerator N*sum(v^2) - sum(v)^2 exactly in integers. The per-row mean and variance stay in double as before. - Add a gtest with the same hand-computed cases as the HiFi test. Reviewed By: DrJessop Differential Revision: D126951724
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:
The generic quantized_layer_norm kernel has the same int16 gap as the HiFi kernel fixed earlier in this stack. The generic kernel is the portable implementation used on CPU and on Xtensa cores that have no backend-specific layer norm kernel.
Before this change, both overloads dispatched with ET_FORALL_CADENCE_QUANTIZED_TYPES, so int16 fell through to an ET_DCHECK_MSG that compiles out in release builds and left the output unwritten. The statistics also used int32 accumulators, which overflow for int16 because one squared value can reach 2^30. The variance was formed as E[x^2] - E[x]^2, which cancels catastrophically when the mean is large compared to the spread.
This applies the same fix as the HiFi kernel:
Reviewed By: DrJessop
Differential Revision: D126951724