Add a16w8 MHA softmax FVP coverage for Ethos-U85 (#19493)#19493
Merged
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/19493
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: ❌ 3 New Failures, 1 Cancelled Job, 1 Pending, 8 Unrelated FailuresAs of commit db06d0b with merge base f1062a7 ( NEW FAILURES - The following jobs have failed:
CANCELLED JOB - The following job was cancelled. Please retry:
FLAKY - The following jobs failed but were likely due to flakiness present on trunk:
BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Contributor
|
@Ninja91 has exported this pull request. If you are a Meta employee, you can view the originating Diff in D103734699. |
This PR needs a
|
Contributor
There was a problem hiding this comment.
Pull request overview
Adds new Arm backend test coverage for torch.softmax in an MHA-like shape under a16w8 quantization, including a sweep over realistic attention-logit ranges and an expected-failure annotation for a known Ethos-U85/Vela numerics issue.
Changes:
- Add
ops/test_softmax.pyto the Arm Bazel test target list. - Introduce an MHA-shaped softmax module and range-sweep INT tests for Ethos-U55 and Ethos-U85 (U85 marked xfail with
strict=False).
Reviewed changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.
| File | Description |
|---|---|
| backends/arm/test/targets.bzl | Adds softmax tests to the Arm test suite target list. |
| backends/arm/test/ops/test_softmax.py | Adds a16w8 MHA-shaped softmax sweep tests for U55/U85, with U85 xfail tied to Vela issue #23. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Ninja91
added
the
partner: arm
Ninja91
force-pushed
the
export-D103734699
branch
from
dc79f31 to
3f89f16
Compare
meta-codesync
Bot
changed the title
Ninja91
added a commit
to Ninja91/executorch
that referenced
this pull request
May 12, 2026
Summary:
Adds a16w8 (int16 IO + int8 weights) coverage for `torch.softmax` in a multi-head-attention shape. Sweeps pre-softmax input ranges to surface a known Ethos-U85 numerics issue: int16 ReduceSum produces silent zero output, which propagates through the standard softmax decomposition (`amax → sub → exp → sum → reciprocal → mul`).
## What's added
`MultiHeadAttentionSoftmax` is a generic MHA-shaped softmax (H=4 heads, M=1 query token, W=16 K/V window). `test_mha_softmax_a16w8_{u55,u85}_INT` sweeps 7 pre-softmax input ranges from `[-0.01, 0.01]` to `[-30, 30]`, covering realistic post-`1/√d` attention logits.
## Tolerances
`atol=0.003` (single value). Calibrated from observed FVP max-abs softmax error at `qtol=0`, sized at ~1.5× the worst-case observed value across the sweep. `rtol` and `qtol` use framework defaults.
## XFAIL handling
The U85 a16w8 cases are wrapped with `pytest.mark.xfail(strict=False)` referencing the upstream Vela report:
https://gitlab.arm.com/artificial-intelligence/ethos-u/ethos-u-vela/-/issues/23
`strict=False` keeps the test target green both on stock Vela 5.0 (cases XFAIL) and once the upstream fix lands (cases XPASS).
Differential Revision: D103734699
Ninja91
force-pushed
the
export-D103734699
branch
from
3f89f16 to
f3370d3
Compare
Ninja91
added a commit
to Ninja91/executorch
that referenced
this pull request
May 12, 2026
Summary:
Adds a16w8 (int16 IO + int8 weights) coverage for `torch.softmax` in a multi-head-attention shape. Sweeps pre-softmax input ranges to surface a known Ethos-U85 numerics issue: int16 ReduceSum produces silent zero output, which propagates through the standard softmax decomposition (`amax → sub → exp → sum → reciprocal → mul`).
## What's added
`MultiHeadAttentionSoftmax` is a generic MHA-shaped softmax (H=4 heads, M=1 query token, W=16 K/V window). `test_mha_softmax_a16w8_{u55,u85}_INT` sweeps 7 pre-softmax input ranges from `[-0.01, 0.01]` to `[-30, 30]`, covering realistic post-`1/√d` attention logits.
## Tolerances
`atol=0.003` (single value). Calibrated from observed FVP max-abs softmax error at `qtol=0`, sized at ~1.5× the worst-case observed value across the sweep. `rtol` and `qtol` use framework defaults.
## XFAIL handling
The U85 a16w8 cases are wrapped with `pytest.mark.xfail(strict=False)` referencing the upstream Vela report:
https://gitlab.arm.com/artificial-intelligence/ethos-u/ethos-u-vela/-/issues/23
`strict=False` keeps the test target green both on stock Vela 5.0 (cases XFAIL) and once the upstream fix lands (cases XPASS).
Differential Revision: D103734699
Ninja91
force-pushed
the
export-D103734699
branch
from
f3370d3 to
7908d4e
Compare
Summary:
Adds a16w8 (int16 IO + int8 weights) coverage for `torch.softmax` in a multi-head-attention shape. Sweeps pre-softmax input ranges to surface a known Ethos-U85 numerics issue: int16 ReduceSum produces silent zero output, which propagates through the standard softmax decomposition (`amax → sub → exp → sum → reciprocal → mul`).
## What's added
`MultiHeadAttentionSoftmax` is a generic MHA-shaped softmax (H=4 heads, M=1 query token, W=16 K/V window). `test_mha_softmax_a16w8_{u55,u85}_INT` sweeps 7 pre-softmax input ranges from `[-0.01, 0.01]` to `[-30, 30]`, covering realistic post-`1/√d` attention logits.
## Tolerances
`atol=0.003` (single value). Calibrated from observed FVP max-abs softmax error at `qtol=0`, sized at ~1.5× the worst-case observed value across the sweep. `rtol` and `qtol` use framework defaults.
## XFAIL handling
The U85 a16w8 cases are wrapped with `pytest.mark.xfail(strict=False)` referencing the upstream Vela report:
https://gitlab.arm.com/artificial-intelligence/ethos-u/ethos-u-vela/-/issues/23
`strict=False` keeps the test target green both on stock Vela 5.0 (cases XFAIL) and once the upstream fix lands (cases XPASS).
Differential Revision: D103734699
Ninja91
force-pushed
the
export-D103734699
branch
from
7908d4e to
db06d0b
Compare
digantdesai
approved these changes
May 12, 2026
Contributor
digantdesai
left a comment
digantdesai
left a comment
There was a problem hiding this comment.
Review automatically exported from Phabricator review in Meta.
usamahz
pushed a commit
to usamahz/executorch
that referenced
this pull request
May 13, 2026
Differential Revision: D103734699 Pull Request resolved: pytorch#19493
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:
Adds a16w8 (int16 IO + int8 weights) coverage for
torch.softmaxin a multi-head-attention shape. Sweeps pre-softmax input ranges to surface a known Ethos-U85 numerics issue: int16 ReduceSum produces silent zero output, which propagates through the standard softmax decomposition (amax → sub → exp → sum → reciprocal → mul).What's added
MultiHeadAttentionSoftmaxis a generic MHA-shaped softmax (H=4 heads, M=1 query token, W=16 K/V window).test_mha_softmax_a16w8_{u55,u85}_INTsweeps 7 pre-softmax input ranges from[-0.01, 0.01]to[-30, 30], covering realistic post-1/√dattention logits.Tolerances
atol=0.003(single value). Calibrated from observed FVP max-abs softmax error atqtol=0, sized at ~1.5× the worst-case observed value across the sweep.rtolandqtoluse framework defaults.XFAIL handling
The U85 a16w8 cases are wrapped with
pytest.mark.xfail(strict=False)referencing the upstream Vela report:https://gitlab.arm.com/artificial-intelligence/ethos-u/ethos-u-vela/-/issues/23
strict=Falsekeeps the test target green both on stock Vela 5.0 (cases XFAIL) and once the upstream fix lands (cases XPASS).Differential Revision: D103734699