Repository navigation
Keep the FACTO pow exponent representable in float32 (#24179) - #24179
Merged
Merged
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24179
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit dfe9115 with merge base bf9bef3 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Contributor
|
@aliafzal has exported this pull request. If you are a Meta employee, you can view the originating Diff in D127746756. |
This PR needs a
|
ethansfng
approved these changes
Oct 27, 2026
DrJessop
approved these changes
Oct 27, 2026
meta-codesync
Bot
changed the title
aliafzal
force-pushed
the
export-D127746756
branch
from
October 28, 2026 06:35
3c82dc8 to
c0dd09e
Compare
aliafzal
added a commit
to aliafzal/executorch
that referenced
this pull request
Oct 28, 2026
Summary: `pow.Tensor_Scalar`'s `exponent` fell into `facto_testcase_gen`'s generic scalar branch, `Gt(0)` and `Le(2)`. The open lower bound makes `nextup(0) = 5e-324` an edge value. That value flushes to 0 when a float32 kernel reads it, so `0 ** 5e-324` (0 in eager, computed with a double exponent) becomes `0 ** 0 = 1` on device. The exponent now gets its own lower bound of `torch.finfo(torch.float32).tiny`. Integer exponents are unchanged (≥ 1), and the upper bound stays 2. Found by the Jarvis nightly operator tests: `test_pow_Tensor_Scalar_out_20` (int32 input, exponent `5e-324`) fails on nxp, DLA, Artemis v3/v4 and CPU with `max_abs_diff = 1.0`. Reviewed By: DrJessop Differential Revision: D127746756
aliafzal
force-pushed
the
export-D127746756
branch
from
October 28, 2026 07:39
c0dd09e to
ab68f86
Compare
aliafzal
added a commit
to aliafzal/executorch
that referenced
this pull request
Oct 28, 2026
Summary: `pow.Tensor_Scalar`'s `exponent` fell into `facto_testcase_gen`'s generic scalar branch, `Gt(0)` and `Le(2)`. The open lower bound makes `nextup(0) = 5e-324` an edge value. That value flushes to 0 when a float32 kernel reads it, so `0 ** 5e-324` (0 in eager, computed with a double exponent) becomes `0 ** 0 = 1` on device. The exponent now gets its own lower bound of `torch.finfo(torch.float32).tiny`. Integer exponents are unchanged (≥ 1), and the upper bound stays 2. Found by the Jarvis nightly operator tests: `test_pow_Tensor_Scalar_out_20` (int32 input, exponent `5e-324`) fails on nxp, DLA, Artemis v3/v4 and CPU with `max_abs_diff = 1.0`. Reviewed By: DrJessop Differential Revision: D127746756
meta-codesync
Bot
changed the title
aliafzal
force-pushed
the
export-D127746756
branch
from
October 28, 2026 20:40
ab68f86 to
a8b84fc
Compare
…torch#24176) Summary: `add.Scalar(Tensor self, Scalar other, Scalar alpha=1)` takes `alpha` positionally, but `add.Tensor(Tensor self, Tensor other, *, Scalar alpha=1)` takes it as keyword-only, and the same holds for `sub`. When `alpha` was the third positional argument, `get_replacement` passed it through `*args[2:]`, and building the `.Tensor` node raised `RuntimeError: aten::add() takes 2 positional argument(s) but 3 was/were given`. The 3-argument path now binds the call by name with `torch.fx.operator_schemas.normalize_function` (the part of `get_arg` that doesn't need a `Node`; `call_operator` gets no node), passes `self` (`input` after normalization) and `other` positionally, and `alpha` as a kwarg. `get_replacement` always gets 2 args and no longer forwards extra ones. The Arm `ReplaceScalarWithTensorByProfilePass` subclass routes through the same base `call_operator`, so it gets the fix too for edge `add/sub.Scalar`. `torch.ops.aten.add/sub.Scalar` with a positional `alpha` still pass through unreplaced, as before. The changed path previously always raised, so no graph that worked before is affected. Authored with Claude Code. Reviewed By: DrJessop Differential Revision: D127746753
…#24177) Summary: `pow.Tensor_Scalar(int32, 2.0)` returns float32, but the `mul.Tensor(x, x)` chain that `ReplacePowWithMulPass` emits stays int32. The pass copied the float32 `meta` onto the int32 `mul` nodes, so the lowered program's output dtype no longer matched the eager reference. The pass now gives up when the input dtype differs from the `pow` output dtype. Authored with Claude Code. Reviewed By: DrJessop Differential Revision: D127746754
…ch#24178) Summary: `div.Tensor_mode(a, b, rounding_mode=None)` is true division. The HiFi `div_out_mode` sent `None` to the non-optimized path, where `mode_is_trunc` was false and the code fell through to `floor_divide` (its comment assumed the mode "must be floor"). So `7 / 2` returned 3 and `-7 / 2` returned -4. The kernel now delegates to `div_out` when `mode` has no value, which is what the portable and Fusion G3 `div_out_mode` already do. Authored with Claude Code. Reviewed By: DrJessop Differential Revision: D127746755
Summary: `pow.Tensor_Scalar`'s `exponent` fell into `facto_testcase_gen`'s generic scalar branch, `Gt(0)` and `Le(2)`. The open lower bound makes `nextup(0) = 5e-324` an edge value. That value flushes to 0 when a float32 kernel reads it, so `0 ** 5e-324` (0 in eager, computed with a double exponent) becomes `0 ** 0 = 1` on device. The exponent now gets its own lower bound of `torch.finfo(torch.float32).tiny`. Integer exponents are unchanged (≥ 1), and the upper bound stays 2. Authored with Claude Code. Reviewed By: DrJessop Differential Revision: D127746756
aliafzal
force-pushed
the
export-D127746756
branch
from
October 28, 2026 20:57
a8b84fc to
dfe9115
Compare
meta-codesync
Bot
changed the title
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:
pow.Tensor_Scalar'sexponentfell intofacto_testcase_gen's generic scalar branch,Gt(0)andLe(2). The open lower bound makesnextup(0) = 5e-324an edge value. That value flushes to 0 when a float32 kernel reads it, so0 ** 5e-324(0 in eager, computed with a double exponent) becomes0 ** 0 = 1on device. The exponent now gets its own lower bound oftorch.finfo(torch.float32).tiny. Integer exponents are unchanged (≥ 1), and the upper bound stays 2.Authored with Claude Code.
Reviewed By: DrJessop
Differential Revision: D127746756