Add a16w8 per-op test for bmm (#19599)#19599
Open
christine-long-meta wants to merge 7 commits into
Open
Conversation
Summary: Add int16 activation / int8 weight (a16w8) quantization tests for `aten.exp` on Ethos-U55 and Ethos-U85. ## Context The `exp` op is part of the softmax decomposition (`softmax(x) = exp(x) / sum(exp(x))`), which is used in the attention mechanism of EMG2Pose Conformer models. This op was identified as the root cause of the U85 SNR regression investigated in SEV T267939669 — without dedicated a16w8 per-op coverage, the numerics issue was only visible at the full-model level. Adding per-op tests allows us to catch int16 precision regressions at the operator granularity before they propagate to end-to-end model accuracy. ## Changes - Add `a16w8_exp_test_parameters` dict with 3 test configurations covering rank-1, rank-2, and rank-3 tensors - Add `test_exp_a16w8_u55_INT` using `EthosU55PipelineINT` with `a16w8_quantization=True, symmetric_io_quantization=True, qtol=128, epsilon=2**-16` - Add `test_exp_a16w8_u85_INT` using `EthosU85PipelineINT` with same kwargs - Register `ops/test_exp.py` in `fbcode/` and `xplat/` `targets.bzl` bypass-pytorch-oss-checks Differential Revision: D104532358
Summary: Add int16 activation / int8 weight (a16w8) quantization tests for `aten.reciprocal` on Ethos-U55 and Ethos-U85. ## Context The `reciprocal` op is the second half of the softmax decomposition (`softmax(x) = exp(x) * reciprocal(sum(exp(x)))`), paired with `exp`. Together they form the attention mechanism in EMG2Pose Conformer models. Like `exp`, this op was implicated in the U85 SNR regression (SEV T267939669) — the division-by-reciprocal path can amplify quantization error when the denominator is itself quantized at int16. Adding dedicated a16w8 coverage isolates reciprocal numerics from the rest of the softmax pipeline. ## Changes - Add `a16w8_reciprocal_test_parameters` dict with 3 test configurations covering rank-1, rank-2, and rank-3 tensors (all shifted by +0.1 to avoid division near zero) - Add `test_reciprocal_a16w8_u55_INT` using `EthosU55PipelineINT` with `a16w8_quantization=True, symmetric_io_quantization=True, qtol=128, epsilon=2**-16` - Add `test_reciprocal_a16w8_u85_INT` using `EthosU85PipelineINT` with same kwargs - Register `ops/test_reciprocal.py` in `fbcode/` and `xplat/` `targets.bzl` bypass-pytorch-oss-checks Differential Revision: D104532357
Summary: Add int16 activation / int8 weight (a16w8) quantization tests for `aten.mean.dim` on Ethos-U55 and Ethos-U85. ## Context The `mean_dim` op is a core component of the LayerNorm decomposition (`LayerNorm = (x - mean) / sqrt(var + eps) * gamma + beta`). It is used across multiple EMG production models including CC, CASCADE, HW, WAKE, and BTD. Despite this wide usage, no a16w8 per-op coverage existed — the int16 quantization path was only exercised indirectly through end-to-end model tests, making it difficult to isolate mean-specific numerics issues from other LayerNorm components. ## Changes - Add `a16w8_mean_test_parameters` dict with 11 test configurations covering keepdim/no-keepdim, positive/negative dims, dim=None, and ranks 1-4 - Add `test_mean_dim_a16w8_u55_INT` using `EthosU55PipelineINT` with `a16w8_quantization=True, symmetric_io_quantization=True, qtol=128, epsilon=2**-16` - Add `test_mean_dim_a16w8_u85_INT` using `EthosU85PipelineINT` with same kwargs - Register `ops/test_mean_dim.py` in `fbcode/` and `xplat/` `targets.bzl` bypass-pytorch-oss-checks Differential Revision: D104532361
Summary: Add int16 activation / int8 weight (a16w8) quantization tests for `aten.var` on Ethos-U55 and Ethos-U85. ## Context The `var` op is the second component of the LayerNorm decomposition, paired with `mean_dim`. Together they compute the normalization statistics used in every LayerNorm layer across EMG models including EMG2Pose. Variance computation is particularly sensitive to int16 quantization because it involves squaring differences — small quantization errors in the mean subtraction are amplified quadratically. Dedicated a16w8 coverage isolates variance numerics from the rest of the LayerNorm pipeline. ## Changes - Add `a16w8_var_test_parameters` dict with 4 test configurations covering keepdim/no-keepdim and correction values 0, 0.5, and 1 - Add `test_var_a16w8_u55_INT` using `EthosU55PipelineINT` with `a16w8_quantization=True, symmetric_io_quantization=True, qtol=128, epsilon=2**-16` - Add `test_var_a16w8_u85_INT` using `EthosU85PipelineINT` with same kwargs - Register `ops/test_var.py` in `fbcode/` and `xplat/` `targets.bzl` bypass-pytorch-oss-checks Differential Revision: D104532362
Summary:
Add int16 activation / int8 weight (a16w8) quantization tests for `aten.conv1d` on Ethos-U55 and Ethos-U85.
## Context
`conv1d` is the most critical op in the EMG stack — it is used by ALL 8 production EMG models (CC, CASCADE, HW, WAKE, BTD, AUTH, EMG2Pose, EMG2Touch) for temporal feature extraction from raw EMG signals. Despite this, only `conv2d` had a16w8 test coverage; `conv1d` was completely uncovered at the int16 activation precision. This gap meant that any Vela or quantizer regression affecting 1D convolutions at int16 IO would go undetected until full-model validation, making root-cause analysis significantly harder.
The test matrix is the largest in this stack because conv1d has the most configuration surface: kernel sizes (1, 3, 5), strides, padding, dilation, depthwise groups, and bias/no-bias variants are all crossed with per-channel vs. per-tensor quantization.
## Changes
- Add `a16w8_conv1d_test_parameters` dict with 14 test configurations (7 conv configs × {per_channel_quant=True, False}) covering kernel sizes 1/3/5, stride 1/2, dilation, depthwise, and no-bias variants
- Add `test_conv1d_a16w8_u55_INT` using `EthosU55PipelineINT` with `a16w8_quantization=True, symmetric_io_quantization=True, per_channel_quantization=<varied>, qtol=128, epsilon=2**-16`
- Add `test_conv1d_a16w8_u85_INT` using `EthosU85PipelineINT` with same kwargs
- Register `ops/test_conv1d.py` in `fbcode/` and `xplat/` `targets.bzl`
bypass-pytorch-oss-checks
Differential Revision: D104532360
Summary: Add int16 activation / int8 weight (a16w8) quantization tests for `aten.gelu` on Ethos-U55 and Ethos-U85. ## Context The `gelu` activation is used in the feed-forward blocks of the EMG2Pose Conformer architecture. GELU is a non-linear activation that involves erf/tanh approximations — these are decomposed into multiple primitive ops during lowering to Ethos-U, making the int16 quantization path particularly susceptible to accumulated rounding error. Without dedicated a16w8 per-op coverage, GELU numerics issues could only be detected at the full Conformer block level. ## Changes - Add `a16w8_gelu_test_parameters` dict with 3 test configurations covering rank-1, rank-2, and rank-3 tensors - Add `test_gelu_a16w8_u55_INT` using `EthosU55PipelineINT` with `a16w8_quantization=True, symmetric_io_quantization=True, qtol=128, epsilon=2**-16` - Add `test_gelu_a16w8_u85_INT` using `EthosU85PipelineINT` with same kwargs - Register `ops/test_gelu.py` in `fbcode/` and `xplat/` `targets.bzl` bypass-pytorch-oss-checks Differential Revision: D104532359
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/19599
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ❌ 1 New Failure, 1 Cancelled JobAs of commit 6a8da35 with merge base 58b4f26 ( NEW FAILURE - The following job has failed:
CANCELLED JOB - The following job was cancelled. Please retry:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
christine-long-meta
force-pushed
the
export-D104532363
branch
from
c0c8f38 to
1854072
Compare
meta-cla
Bot
added
the
CLA Signed
|
|
This PR needs a
|
christine-long-meta
added a commit
to christine-long-meta/executorch
that referenced
this pull request
May 14, 2026
Summary: Add int16 activation / int8 weight (a16w8) quantization tests for `aten.bmm` on Ethos-U55 and Ethos-U85. ## Changes - Add `a16w8_bmm_test_parameters` dict with 5 test configurations covering same-shape, different-shape, rectangular, batch-10, and negative-value tensors - Add `test_bmm_a16w8_u55_INT` using `EthosU55PipelineINT` with `a16w8_quantization=True, symmetric_io_quantization=True, qtol=128, epsilon=2**-16` - Add `test_bmm_a16w8_u85_INT` using `EthosU85PipelineINT` with same kwargs - Remove unused `aten_op_mm` and `exir_op_mm` variables - Register `ops/test_bmm.py` in `fbcode/` and `xplat/` `targets.bzl` bypass-pytorch-oss-checks Differential Revision: D104532363
christine-long-meta
force-pushed
the
export-D104532363
branch
from
1854072 to
d5349b3
Compare
meta-codesync
Bot
changed the title
christine-long-meta
force-pushed
the
export-D104532363
branch
from
d5349b3 to
77a9a41
Compare
Contributor
|
@christine-long-meta has exported this pull request. If you are a Meta employee, you can view the originating Diff in D104532363. |
1 similar comment
Contributor
|
@christine-long-meta has exported this pull request. If you are a Meta employee, you can view the originating Diff in D104532363. |
Summary: Pull Request resolved: pytorch#19599 Add int16 activation / int8 weight (a16w8) quantization tests for `aten.bmm` on Ethos-U55 and Ethos-U85. ## Changes - Add `a16w8_bmm_test_parameters` dict with 5 test configurations covering same-shape, different-shape, rectangular, batch-10, and negative-value tensors - Add `test_bmm_a16w8_u55_INT` using `EthosU55PipelineINT` with `a16w8_quantization=True, symmetric_io_quantization=True, qtol=128, epsilon=2**-16` - Add `test_bmm_a16w8_u85_INT` using `EthosU85PipelineINT` with same kwargs - Remove unused `aten_op_mm` and `exir_op_mm` variables - Register `ops/test_bmm.py` in `fbcode/` and `xplat/` `targets.bzl` bypass-pytorch-oss-checks Differential Revision: D104532363
christine-long-meta
force-pushed
the
export-D104532363
branch
from
77a9a41 to
6a8da35
Compare
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
1 participant
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:
Add int16 activation / int8 weight (a16w8) quantization tests for
aten.bmmon Ethos-U55 and Ethos-U85.Changes
a16w8_bmm_test_parametersdict with 5 test configurations covering same-shape, different-shape, rectangular, batch-10, and negative-value tensorstest_bmm_a16w8_u55_INTusingEthosU55PipelineINTwitha16w8_quantization=True, symmetric_io_quantization=True, qtol=128, epsilon=2**-16test_bmm_a16w8_u85_INTusingEthosU85PipelineINTwith same kwargsaten_op_mmandexir_op_mmvariablesops/test_bmm.pyinfbcode/andxplat/targets.bzlbypass-pytorch-oss-checks
Differential Revision: D104532363